The arithmetic seems neat. The conclusion is often wrong.

The standard version goes something like this. Assume the time spent on a paper is perhaps 2-3 hours, 3-5 papers per year per academic, multiply by salary costs, scale nationally, and arrive at a figure large enough to generate social media outrage. It has the appealing quality of all back-of-the-envelope calculations – it looks rigorous while quietly ignoring everything inconvenient.

Who is actually paying?

The most fundamental problem with the calculation is not the arithmetic. It is the assumption buried inside it: that the university is paying for peer review.

In most cases, it is not.

Peer review is a voluntary activity. No institution I am aware of formally requires academics to undertake it as part of their contractual duties. It does not appear on the list of things a university actively enforces, monitors or makes explicit room for in a workload model. If I take on a review, I do so on top of everything else – and everything else takes priority. The university is not, in any meaningful sense, funding those hours. I am.

This matters enormously, because it completely changes what the calculation is actually measuring. It is not university expenditure. It is academic goodwill – which is a rather different thing, and not something that appears on any institutional balance sheet.

What universities do spend time on

It is worth pausing here to consider what academic time universities do explicitly require and enforce. Meetings, for instance – the kind that arrive in your calendar with the inevitability of rain. Hours of meeting time with a lot of relatively highly-paid people and take home messages that often fill five to ten minutes. Administrative processes that exist because they have always existed. Tasks that cry out for professional support but are assigned to academics on the grounds that they fall, technically, within academic responsibility.

I once spent a considerable stretch of time manually assessing 150 multiple choice question cards because the optical reader was unavailable. There were only 12 questions to answer. Still, this meant 1800 I needed to manually assess. It certainly was not a high point of my scholarly career. Nobody calculated the cost of this undertaking and published it on social media.

The contrast is instructive. Time spent on tasks of genuine intellectual value, chosen voluntarily, goes uncounted. Time spent on enforced administrative process is also uncounted – but for quite different reasons.

Why academics actually do it

If peer review is voluntary and unrewarded institutionally, the obvious question is: why do academics do it at all?

The answer, at least in my experience and that of most colleagues I have spoken to, has very little to do with obligation and rather a lot to do with self-interest – in the best possible sense. Reviewing a manuscript means reading work that has not yet been published. It means spending focused time with new results, new methods and new interpretations from your specific corner of the field, before almost anyone else has seen them. It is, in a discipline where keeping up with the literature is a constant and losing battle, one of the few reliable ways to stay genuinely current.

There is also something to be said for the quality of reading that review demands. A paper you are asked to evaluate properly gets read in a way that published papers, in the relentless churn of academic life, often do not. That enforced close reading has value – for the reviewer as much as for the author.

Add to this the less self-interested reasons: improving the work of colleagues, catching errors before they become part of the permanent record, contributing to a system from which every publishing academic benefits. These are real motivations, and they operate entirely outside any institutional incentive structure.

The reciprocity that the calculation ignores

There is a broader point here about how science functions as a collective enterprise. Every paper that gets published has been reviewed by someone, voluntarily, as part of an informal but remarkably durable system of scholarly reciprocity. If you publish, you benefit from that system. Participating in it is not exploitation – it is the mechanism by which the whole thing works.

The "billions wasted" framing treats peer review as overhead: a drag on productivity, an inefficiency to be optimised away. But that misunderstands what it is. Peer review is not the cost of doing science. It is part of doing science – the part where ideas are tested, challenged and refined by people who know the field well enough to do so usefully.

The real inefficiencies in the system lie elsewhere: in the volume of low-quality submissions, in review requests sent to people with no relevant expertise, in the resubmission cascades that multiply effort without improving work. Those are legitimate problems. But they are problems with how the system is managed, not with the principle of peer review itself.

A note of honesty

I should acknowledge that this view is shaped by my own field and my own institution, and that experiences vary considerably across disciplines and career stages. Colleagues under greater pressure, in fields with different review cultures, may find the balance of costs and benefits looks rather different. The argument for treating peer review as professional development rather than unpaid labour is easier to make from some positions than others.

But the basic point stands: a calculation that assigns the full salary cost of peer review hours to the university, while ignoring that those hours are voluntary, unrequired and often undertaken outside normal working time, is not measuring what it claims to measure. It is producing a large number and calling it an insight.

Science is expensive. Careful scrutiny of new work is part of that cost, and a worthwhile one. We just had the pleasure of having another paper accepted in Nucleic Acids Research after quite a long and detailed peer review process. I have said it before, and I say it again: it felt painful, but the process has significantly strengthened the paper (more on this paper very soon).

Pretending that peer review is simply institutional waste misunderstands both how universities work and what academic research actually is.

The arithmetic, as I said, is often neat.

Similar and related blog posts:

• When the editor get's it wrong
• Termination of DNA replication triggers genomic instability, a paper significantly strengthened by the peer review process
• Some insights into the statistics and scale of the peer review process

In metric, one milliliter of water occupies one cubic centimeter, weighs one gram, and requires one calorie of energy to heat up by one degree centigrade—which is 1 percent of the difference between its freezing point and its boiling point. An amount of hydrogen weighing the same amount has exactly one mole of atoms in it.

Whereas in the American system, the answer to “How much energy does it take to boil a room-temperature gallon of water?” is “Go fuck yourself,” because you can’t directly relate any of those quantities.

The system that sas served us well

A few years ago, after considerable searching, we settled on a temperature monitoring solution for the lab freezers: Testo Saveris 2 data loggers, connected via WiFi to a cloud service that records temperatures continuously and raises alerts when critical limits are breached. The system covers three freezers across two rooms, has caught most real temperature excursion we have had, and has been running reliably for years.

The WiFi network the loggers connect to is not the university network – getting IT to cooperate with that particular request proved fruitless – but a dedicated personal hotspot running on a pay-as-you-go SIM from Three. The data loggers generate almost no data traffic, so this worked out to roughly £5 every two or three years. Simple, cheap, effective.

Or so it seemed.

The gradual accumulation of irritations

The setup had one persistent flaw. Our freezers are split across two rooms: two in a storage room, one in the lab. The router lived in the storage room – logical enough, since that is where the strain collection lives and where WiFi coverage mattered most. But the logger in the lab, further from the router, would occasionally fail to connect. Not catastrophically – it would pick up the connection again within an hour or two – but enough to generate a steady trickle of "connection failed" and "resuming normal function" emails. A minor irritation, nothing more.

The bigger irritations came later, and they came from Three.

First, Three changed their data policy. The days of purchasing a data package that rolled over indefinitely into the next month were over. Suddenly, I was buying a £5 – later £10 – data package every single month, for a system that was using almost none of it. Not a fortune, but deeply unnecessary.

Then came multi-factor authentication. Logging in to purchase that monthly data package now required entering a one-time code sent to the number associated with the SIM. Fair enough in principle. In practice, this meant logging into the router's interface via my phone to retrieve the SMS. Mildly annoying, but manageable.

Then the router's web interface stopped accepting login credentials. Most likely a JavaScript compatibility issue as browsers updated – though I am open to correction on the precise cause. Whatever the reason, I was effectively locked out of the D-Link DWR-932 router that had been quietly doing its job for years. The workaround? Every month: remove the SIM from the router, insert it into my phone, receive the authentication code, purchase the data package, remove the SIM from the phone, reinsert it into the router.

Every. Single. Month.

I contacted Three to ask for alternatives – a SIM with a different data structure, MFA via email rather than SMS, anything that might simplify the situation. Their response was consistent and magnificently unhelpful: had I considered switching to a plan with unlimited texts, unlimited UK calls, and several gigabytes of data?

Anyone who engages their brain for even one second will immediately grasp that unlimited texts and UK calls are of absolutely no relevance whatsoever to a temperature logger in a freezer room. But Three, bless them, remained convinced that more data was always the answer, regardless of the question.

AI to the rescue

This is where things improved considerably – and where AI deserves genuine credit.

The more elegant solution, though, was the SIM. I described the problem to AI – low data volumes, long time horizons, no need for voice or SMS, just reliable connectivity for a handful of small devices – and the answer came back almost immediately: an IoT SIM.

IoT stands for Internet of Things, the catch-all term for the growing ecosystem of devices that connect to the internet without any human directly operating them. Smart meters, environmental sensors, connected equipment – and, as it turns out, laboratory temperature loggers. IoT SIMs are specifically designed for exactly this use case: low data consumption over extended periods, without the monthly package structure of conventional consumer SIMs. They are also not locked to a single network, registering instead with whatever network is available – a very useful bonus.

After a little research I found infiSIM. The deal: £15 for 500 MB of data, valid for five years or until the data runs out – whichever comes first – with a top-up option if needed. Delivery added another £15 for a courier, which stung slightly, but the arithmetic is hard to argue with. Roughly £30 upfront and perhaps a top-up sometime in the next five years, versus £10 a month and a monthly SIM-swapping ritual.

There was one small snag: getting the correct APN settings configured in the new router for the IoT SIM required navigating some menus that were not immediately obvious. AI walked me through the process step by step in plain language, the settings were entered, and the connection was established.

The current state of affairs

The system now looks like this: three Testo Saveris 2 data loggers, connected via a TP-Link TL-MR100 router running an infiSIM IoT SIM, feeding data to the Testo cloud service that monitors temperatures and raises alerts as needed.

The monthly SIM-swapping ritual is gone. The monthly data purchases are gone. The Three customer service conversations are, mercifully, also gone.

The connection issue with the more distant logger in the lab has not been entirely eliminated – the physical distance and intervening walls are what they are – but it occurs roughly ten times less frequently than before. The stronger router has made a substantial difference. The occasional connection dropout still happens, but it has gone from a regular minor irritation to a rare event that is easy to keep an eye on.

With a reasonable degree of optimism, this setup should now run without significant intervention for the next five years. For a system whose entire purpose is to sit quietly in the background and raise an alarm when something goes wrong, "running without significant intervention" is exactly the right outcome.

Sometimes the best technology is the kind you completely forget about.

Equipment mentioned in this post: Testo Saveris 2 data loggers, TP-Link TL-MR100 4G LTE Router, infiSIM IoT SIM.

Similar and related posts:

• Temperature monitoring of freezers
• Temperature profile of an underbench freezer during a power cut
• Temperature profile of a tall upright freezer during a power cut

The problem

This website has been running on Kirby 2.5 for years. Kirby is a flat-file content management system (CMS) – no database, clean and elegant – and version 2.5 served us well for a long time. But software ages, and Kirby 2.5 was beginning to show its years. PHP support for older versions creates compatibility headaches as server environments move forward, and it was becoming increasingly clear that an upgrade to Kirby 5 was not just desirable but necessary.

There was just one problem: upgrading from Kirby 2.5 to Kirby 5 is not the kind of upgrade where you download a new version, click install and carry on with your day. It is, to put it plainly, a migration – and a substantial one.

What a Kirby 2.5 to Kirby 5 Migration Actually Involves

For those unfamiliar with what this kind of upgrade entails, it is worth briefly explaining why it is non-trivial.

Kirby 5 is a fundamentally more modern piece of software than version 2.5. The entire folder and file structure has changed. In Kirby 2.5, content, templates and configuration were organised in ways that simply do not map directly onto the Kirby 5 architecture. Template files need to be rewritten. The snippet system works differently. The Panel – Kirby's control interface – has been rebuilt. Blueprint files, which define how content is structured and how the Panel presents it to the user, use a different syntax and logic. Plugins from the 2.5 era are not compatible and need to be replaced or rebuilt.

In short: you cannot lift the old site and drop it into the new version. You rebuild around the content, carrying forward what matters and rewriting the machinery that drives it.

So far I had nothing to do with any of this, so perhaps understandably it was a very daunting prospect.

The license question

Before touching a single file, there was a practical question to resolve: licensing. Kirby 5 requires a new license, and while Kirby offers reduced pricing for educational use, my situation was not straightforward. I am an academic, but this is not a university website. It is my personal lab website – a place for sharing research, writing, and the kind of content you are reading now. Did that qualify?

Rather than assuming either way, I did what turned out to be the right thing: I simply emailed the Kirby team, explained who I was, what the site was for, and asked.

The response arrived the next day. They were happy to give me a Kirby 5 license.

For free.

I will admit that I sat with that email for a moment before responding. A single license is not enormously expensive, so this was not a dramatic gesture as such. But that is almost the point. It was a small thing for them that landed as a genuinely generous act. No interrogation of my credentials, no lengthy back-and-forth. Just: here is your license, good luck with the migration.

For context, in case you have not read the FileMaker blog post: a few months earlier, a sales representative at Claris had failed to mention that a new version of FileMaker Pro was weeks away from release – information that would have directly changed a purchasing decision worth considerably more than a Kirby license. The contrast could not be sharper.

This is how customer support should work. A huge thank you to the Kirby team for their generosity and for restoring some faith in the idea that software companies can still be run by humans who care about their customers.

The migration: enter Claude Code

With the license sorted, the migration itself still needed to happen. This is where my friend Tim came in – a software engineer who volunteered to help, and whose approach to the problem introduced me to something I had not seen in action before.

Tim's method was to use Claude Code, Anthropic's AI coding assistant, to do the heavy lifting. The process was straightforward in concept but remarkable to watch in practice: the entire site content and codebase was copied to a local folder, and Claude Code was then given access to everything it needed – all the files, all the structure, all the content – through nothing more than plain language instructions.

What struck me was how natural the interaction was. There was little need to specify file paths or explain the architecture in technical terms. You could simply describe what needed to happen – "headings in section XYZ are not rendering correctly", "the blog post template needs to pull the featured image from the page files", "for the main page the panel for uploading images does not show" – and Claude Code would locate the relevant files, understand the context, propose a fix, and implement it. When something did not work as expected, you described the problem in the same plain language and it adjusted. Not always perfectly clean, not always without keeping some legacy coding, but still: after a few hours of work, mostly prompting in AI, the entire content was largely migrated.

For someone who has spent years feeling vaguely intimidated by anything beyond the CMS interface, watching a complex migration handled through what amounted to a conversation was genuinely eye-opening. It did require Tim's expertise to guide the process. There were moments where judgment and experience mattered and where Claude Code's suggestions needed steering. But the volume of work that was handled automatically, correctly and quickly was striking. It would have taken me months to do the same thing myself.

What followed in the next week or two was that I meticulously overhauled every section of the site, identifying further errors and coding issues that were fixed. Content was checked and updated, templates were tested, the Panel was reconfigured. What might have taken months of painstaking manual work was compressed into something far more manageable.

The result

The site you are reading this on is the result. As far as I can tell, it works – and if you notice anything that does not, the contact page is there for a reason.

Kirby 5 is a genuinely excellent CMS. It is clean, fast, flexible and a pleasure to work with. The Panel is well-designed, the content management workflow is intuitive, and the flat-file architecture means there is no database to worry about. If you are considering a CMS for a personal or small organisational site and want something that gets out of your way and lets you focus on content, I would recommend it without hesitation.

And if you contact their support team with an awkward question, may that be about licensing or other issues, I am quite confident that they will surprise you with their knowledge, expertise and generosity.

The FileMaker Pro experience referenced in this post is documented in an earlier entry on this blog.

The email that AI handled better than I could

Not long ago I found myself needing to write a difficult email. The recipient would not welcome its contents. It had to be transparent, precise and fully compliant with university rules and regulations. As a mediator in the situation, I also needed to ensure that nothing in the message could leave me exposed to complaints. I was, in essence, only the messenger – the message had to be correct, clear and carefully worded.

The scenario practically demanded AI assistance. So I used one of the available platforms. I explained the situation, the required outcome, the regulatory constraints, the tone. The process took perhaps ten minutes of careful explanation on my part. Then, in seconds, the platform generated the message.

It was exactly right. Clear, compliant, professionally phrased and constructed in a way that protected my position entirely. If I am honest, it was better than what I would have produced myself. In that context, AI was not just useful – it was genuinely valuable in a way that is hard to argue with.

The second email

The same day, I needed to write another message. It was less complex, closer to the kind of communication I handle routinely as part of my daily work. There was no particular regulatory minefield to navigate, no unusual sensitivity to manage. An ordinary, professional email.

And I simply could not be bothered to write it myself.

Why engage my brain when I could sketch out a few points, hand them to AI and receive a polished, professional message within seconds? The elegant solution from earlier in the day had reset my expectations. The bar for what felt worth doing manually had shifted, almost without my noticing.

This is where the problem begins. The temptation is large. And it comes at a cost.

Enter the research

A recently published preprint provides causal evidence that what I experienced is not merely personal idiosyncrasy. The paper, by researchers at Oxford, MIT and UCLA, reports findings from a series of randomised controlled trials involving over 1200 participants. Across tasks including mathematical reasoning and reading comprehension, they found that while AI assistance improved performance in the short term, participants who had used AI performed significantly worse on subsequent unassisted tasks. Importantly, they were more likely to give up. These are short, controlled tasks, so the results reflect immediate behavioural changes rather than long-term cognitive decline.

What makes this particularly striking is the speed at which the effects emerged. Measurable reductions in persistence and independent performance appeared after approximately 10 min of AI interaction. Not months of habitual use. Ten minutes!

The authors argue that AI conditions people to expect immediate answers, reducing their tolerance for productive difficulty – the kind of struggle that is not incidental to learning, but central to it. A mentor or tutor scaffolds learning precisely by not always providing the answer. Current AI systems are optimised to do the opposite. What this suggests is not just a performance effect, but a shift in effort strategy: once an answer is available, people invest less in solving the problem themselves.

A Note on the research itself

Before going further, the limitations of this study are worth stating clearly. It is a preprint, currently under review, and has not yet been independently replicated. The tasks examined – structured mathematical and reading problems – are well-suited to controlled experimentation, but are narrower than the full range of ways people use AI in real professional and academic contexts. There is no preregistration or open data currently available, which limits the ability to audit the findings independently.

But, with these caveats clearly in place: the findings resonate.

What practice actually does

Some years ago I had a productive academic year in which I wrote four research papers in succession. The first was painful from start to finish. The content was demanding, but so was the writing itself. It had been some time since I had written a full paper, and it showed. Multiple rounds of revision were required before it was finally accepted for publication.

The second paper was easier. The third, easier still. By the fourth, the manuscript came together in roughly two weeks. Co-authors required minimal revisions. Peer review went smoothly. The content of that paper was no less complex than the first – perhaps even more complex – but the continuous practice had transformed the process.

Writing, it turns out, is a skill that responds to training in the way most skills do. The more you do it, the more fluent it becomes. The connections between thinking and expression grow stronger. The ability to structure an argument, anticipate a reader's objections and calibrate tone all improve – but only through repeated, effortful practice.

AI, used unreflectively, removes exactly that practice.

The real question

None of this means AI has no place in academic or professional writing. The complex email I described earlier is a genuinely good use case. When a task requires holding in mind a large network of constraints – regulatory requirements, institutional tone, professional risk, precise language – AI's ability to manage that complexity simultaneously is remarkable. It does something that is genuinely difficult to do well unassisted.

The question is not whether to use AI. The question is how.

The risk is not the tool itself but the habit of reaching for it automatically, including for tasks that would benefit from the effort of doing them yourself. The professor who has had an assistant handle everything for so long that they can no longer draft a straightforward memo is not a hypothetical figure. That trajectory is now available to anyone with a smartphone, tablet of computer to access AI – and it can apparently begin within ten minutes.

What the preprint suggests, and what my own experience supports, is that the cost of AI convenience is not always visible at the moment of use. It accumulates quietly, in the form of skills not practised, difficulties not worked through, and a gradually lowered threshold for engaging with hard problems independently.

A more useful relationship with AI

The more productive approach – easier to describe than to maintain – is to use AI as a tool for learning rather than a replacement for thinking. When AI produces a well-structured email or a clearly argued paragraph, the response that serves you best is not simply to send it, but to read it carefully and understand why it works. What did it do that you did not? What did it see that you missed? How is it constructed? In fact, current tools even provide the info: often you get sections such as "why this works better", allowing exactly this type of analysis.

Used that way, AI becomes something closer to the mentor the preprint describes: something that scaffolds your development rather than substituting for it. The output is still useful. But the process also makes you better.

That requires discipline, particularly when the alternative – sketch a few points, receive a polished result, move on – is frictionless. The temptation is real. I know this because I gave in to it on an ordinary Tuesday afternoon, writing a perfectly routine email.

The research suggests I was not alone. And that the consequences may be more significant than a single email implies. The question is no longer whether AI helps, but what it changes in how we approach thinking itself.

Related and similar blogposts:

• Scientific AI platform that is genuinely useful
• LINER AI peer review
• When AI cannot count
• The free Scite.ai browser plugin

What it actually does

AI2 Scholar is an AI-powered research assistant developed by the Allen Institute for Artificial Intelligence. Unlike general AI tools, it is built explicitly for scientific workflows, searching across over 100 million abstracts and millions of full-text papers to help researchers navigate academic literature.

The key difference from tools like ChatGPT or Claude: it does not try to answer questions from its training data. Instead, it searches scientific literature, synthesizes information from multiple papers and provides evidence-based answers grounded in cited sources.

For reliability this matters.

How it works in practice

AI2 Scholar is not designed as a conversational interface. You cannot have a back-and-forth chat to refine your query. Instead, you need to formulate a clear, focused question that can be answered by published research. This requires some thought upfront – you are investing time in crafting a good query rather than iterating through dialogue.

Submit a well-formed question and you receive a structured report, typically organized with an introduction followed by multiple subsections, each addressing specific aspects of your query. The introduction is usually AI-generated summary without direct citations, but the subsequent sections are referenced.

Here is where it gets particularly useful: every reference is clickable. Click on a citation and a pop-up window appears showing not just the paper title and authors, but the exact section AI2 Scholar used to support its claim. You can read the precise text it is citing.

This transparency reveals something important that anyone who hass chased references knows well: claims followed immediately by citations often mean the cited source is not actually the original – it is citing someone else. With AI2 Scholar, you can spot this relatively quickly. The system points you toward relevant literature, and you can quickly identify which sources need further investigation to find the actual original work.

This is enormously valuable. It does not eliminate the work of proper literature review, but it focuses your effort where it matters: verifying claims, finding original sources, evaluating evidence. The semi-automated approach means you are spending time on critical thinking and verification rather than initial literature trawling.

Real-world applications

I have used AI2 Scholar across a surprising range of scientific topics. When preparing exam questions about human physiology – outside my primary research area – it provided rapid, well-referenced overviews that helped me frame questions accurately. For research paper preparation, it has delivered focused summaries of bacterial physiology topics I am less familiar with, complete with citations, some of which I would have missed. The clickable references let me immediately assess which sources were reviews versus primary research, which claims needed verification and where the field's current understanding stood. It provides a very solid base to build upon.

Or when reading a paper that uses the PURE transcription-translation system. This cell-free protein synthesis system involves considerable biochemistry outside my immediate expertise. A focused query to AI2 Scholar generated a structured report covering mechanism, applications and limitations, each section citing relevant papers.

Could I have found this information through traditional literature searches? Of course, absolutely. But it would have taken substantially longer and I would have missed some of the important sources. The value is not that AI2 Scholar finds papers I could not locate myself. It is that it assembles a coherent overview from multiple sources and lets me quickly evaluate source quality.

The "Find Papers" function

Beyond generating reports, AI2 Scholar's paper search function accepts natural language queries describing what you are actually looking for. Rather than constructing Boolean keyword searches for PubMed, you can simply describe the research question.

This often produces more valuable results than traditional keyword approaches. Keyword searches require knowing the precise terminology used in a field, which becomes circular when you are entering unfamiliar territory. Describing what you want to know in plain language frequently surfaces relevant papers that keyword searches would miss because they use different terminology or frame the question from a different angle.

The system understands context and relationships between concepts in ways that keyword matching simply does not. It is not magic – it is large-scale language models applied to scientific literature. But the practical effect is noticeably better literature discovery.

What it is not for

AI2 Scholar is deliberately narrow in scope. It will not help with:

General questions unrelated to scientific literature. It searches academic publications, not the broader internet or general knowledge.

Tasks requiring back-and-forth refinement of ideas. The lack of conversational interface means you need clarity upfront rather than iterative exploration.

Non-research applications like coursework marking, code generation, or creative writing. Other AI tools serve these purposes better.

This specificity is not a limitation, it clearly is a feature, and in my opinion an excellent one. By focusing exclusively on scientific literature synthesis, the tool does that one thing well, better than other more flexible systems.

Why students should use this

I wish more students would discover AI2 Scholar for coursework research. The quality of literature reviews in student assignments would improve if they used this instead of asking more commonly used tools such as ChatGPT, Claude or Gemini.

AI2 Scholar forces good practices: formulate a clear question, examine actual sources, verify citations. It provides scaffolding for literature review without doing the thinking for you. The clickable references mean students can – and must – engage with primary literature rather than accepting AI-generated summaries uncritically.

The reports it generates are still just starting points. You need to read the sources, evaluate their quality, understand their methods, assess their conclusions. But starting from a well-organized, properly cited overview is vastly superior to either traditional keyword searching or asking general AI tools for answers they will hallucinate from training data.

AI2 Scholar handles the first part. The author has to handle the second. The combination is substantially more efficient than either alone.

One more thing

It is completely free to use.

For a tool this useful – no subscription fees, no usage limits, no upselling to premium features. That alone makes it worth trying. The Allen Institute for AI operates as a non-profit research organization with a focus on open science. They are not monetizing your searches or limiting functionality behind paywalls.

If you regularly work with scientific literature, whether for research, teaching or learning, AI2 Scholar is worth exploring. It will not replace your expertise or eliminate the work of proper literature review. But it might focus that work more effectively and save substantial time in the process.

I have briefly tried many different LLM-based platforms, often without ever returning. I consistently find myself returning to AI2 Scholar.

Similar and related blog posts:

• LINER peer review
• Navigating the literature landscape
• Paper-Wizard review tool
• The Scite.ai browser plugin

The answer is not as obvious as it seems. When you consider what freezing actually does to living cells – forming ice crystals that physically damage membranes, creating severe osmotic stress, triggering recrystallization during thawing – it is remarkable that anything survives at all. Yet we routinely freeze bacterial cultures at –80°C or in liquid nitrogen at –196°C, store them for years and revive them months or even decades later with a relatively moderate loss of viability.

The accidental experiment

A student in our lab recently demonstrated this resilience quite unintentionally. They froze a bacterial stock with only 5% glycerol instead of the standard 40% that we use. Even though our concentration is high, 5% is definitely well below the recommended concentration for cryopreservation. When I discovered the error six months later and pulled the stock from the freezer, I naively expected the stock to be dead.

However, the culture grew perfectly fine. In fact, as robustly as one of my stocks I generated at the same time with the correct glycerol concentration.

This should not necessarily be a surprise, but it does highlight something important: bacterial cells possess remarkable natural tolerance to freezing damage. The glycerol concentrations we use in standard protocols are not strictly necessary for survival—they're there to maximize it.

What freezing actually does

To understand why bacteria survive freezing, we need to understand what makes freezing so destructive.

As water freezes, ice crystals begin forming in the extracellular medium. These growing crystals exclude solutes into an ever-decreasing volume of liquid water, creating severe osmotic stress. Cells experience dramatic increases in salt concentration in the remaining unfrozen water, which can denature proteins and disrupt membrane integrity. If you have ever generated a concentration gradient by freezing, you will recognise this effect immediately. Incidentally, this also is the explanation why frozen samples, if thawed, always should be mixed thoroughly.

Intracellular ice formation represents an even more direct threat. Ice crystals forming inside cells can physically puncture membranes, create gas bubbles and disrupt cellular structures. The mechanical damage alone should be lethal.

Then there is the thawing process. If warming occurs too slowly, small ice crystals recrystallize into fewer but larger crystals, which proves rather destructive. The period just before the melting point – when ice growth and recrystallization occur most rapidly – is especially dangerous.

Given all this, how do bacterial cells survive?

Natural cryoprotection: bacteria come prepared

Bacteria have evolved in environments where temperature fluctuations are common. Many species regularly encounter freezing conditions and have developed sophisticated protection mechanisms that help them endure extreme cold.

One example (among several) is Polyhydroxybutyrate (PHB) granules, which represents one of the most effective bacterial cryoprotection strategies. These intracellular storage polymers retain remarkable flexibility even at extremely low temperatures. PHB granules protect cells by maintaining membrane flexibility and facilitating higher rates of transmembrane water transport, which prevents the formation of lethal intracellular ice crystals. The protective effect becomes more pronounced at lower temperatures, precisely when cells need it most.

Another example is antifreeze proteins, which minimize freezing damage by inhibiting the growth of large ice crystals. These proteins are found across numerous bacterial genera and protect cellular structures during both freezing and thawing. Rather than preventing ice formation entirely, they control crystal size and growth rates, limiting mechanical damage.

There are numerous other strategies as well.

The single-cell advantage

Perhaps the most important factor in bacterial freezing tolerance is structural simplicity. Bacteria are single-celled organisms with relatively simple organization. They do not have delicate tissue structures, complex organ systems, or intricate cellular networks that can be disrupted by ice formation.

When bacterial cells freeze, each cell is an independent unit. Some cells in a population will die: ice crystal formation and osmotic stress will damage membranes beyond repair in a fraction of the culture. But the survivors remain fully functional. There is no tissue architecture to maintain, no blood vessels to rupture, no nerve connections to preserve. A bacterial cell either survives intact or it does not, and the survivors can immediately resume normal growth upon thawing.

This explains why even suboptimal freezing (like our student's 5% glycerol stock) still works. Yes, more cells die than would with proper cryoprotectant concentrations. But enough survive to establish a viable culture, and those survivors are entirely normal.

Why we use proper protocols anyway

If bacteria can survive freezing with minimal protection, why do we bother with 15–40% glycerol, controlled cooling rates and rapid thawing protocols?

The answer is maximizing survival.

Standard cryopreservation protocols dramatically improve the proportion of cells that survive freezing. Appropriate cooling rates and rapid freezing approaches (flash freezing) help minimise intracellular ice formation while minimizing damage from excessive dehydration. Cryoprotectants like dimethyl sulfoxide (DMSO) and glycerol bind intracellular water, preventing ice crystal formation and reducing salt concentration in the remaining liquid. Rapid thawing minimizes the dangerous recrystallization period when small ice crystals merge into larger, more destructive ones.

These optimizations matter for practical applications. Higher survival rates mean more reliable stock revival, better experimental reproducibility and longer viable storage times. Some bacteria stored properly in liquid nitrogen remain viable for decades.

Why this does not work for complex organisms

The ease of bacterial cryopreservation becomes more remarkable when compared to the difficulties faced by researchers working with other model organisms. It was only when I was working as a postdoc and interacted with a number of colleagues working with other models that I realised that zebrafish researchers cannot freeze adult fish. They must maintain breeding colonies continuously. Drosophila geneticists face the same constraint with fly lines. C. elegans laboratories have only recently achieved success freezing embryos, and even that remains technically demanding. Adult worms and larvae cannot be frozen at all.

This creates substantial logistical challenges. Maintaining hundreds of genetic lines through continuous breeding requires significant space, resources and labor. A single freezer failure for bacterial stocks is inconvenient; losing a unique Drosophila line means it is gone permanently unless maintained elsewhere.

Why can we not freeze these organisms when bacteria freeze so readily?

The problems are multiple. First, tissue architecture cannot survive the freezing process even if individual cells do. Blood vessels rupture, nerve connections break, organ structures collapse. Even if every cell survived individually – which they do not – the organism would be destroyed by loss of tissue organization.

Second, cryoprotectants must penetrate every cell in a complex organism. This is slow, toxic at necessary concentrations and fundamentally uneven. In organisms with exoskeletons or other barriers, penetration may be impossible at physiologically tolerable concentrations.

Third, size creates uniformity problems. Larger organisms cannot freeze homogeneously. Outer layers freeze first, inner tissues later, creating gradients in ice crystal size, osmotic stress and recrystallization dynamics that prove impossible to control during thawing.

Bacteria avoid all these problems. As single cells, they have no tissue architecture to maintain. Their small size allows rapid, uniform freezing and thawing. Cryoprotectants can reach the entire cell easily. And their natural protection mechanisms – evolved over billions of years encountering environmental freezing – provide baseline resilience that complex organisms simply do not possess.

Appreciating the convenience

The next time you pull a bacterial stock from the freezer and watch colonies grow on a plate the following day, it is worth appreciating what is actually happening. Those cells survived freezing at –80 °C. A process that forms ice crystals, creates severe osmotic stress and generates oxidative damage during thawing. They did so partly because of the glycerol we added, perhaps the controlled freezing rate if used, and the rapid thaw protocol we followed.

But they also survived because they are bacteria: single-celled organisms with natural cryoprotection mechanisms, flexible membranes and billions of years of evolutionary experience with environmental temperature extremes. We have optimized the process, but bacteria made it possible.

Meanwhile, our colleagues working with flies, fish and worms maintain continuous breeding colonies because their organisms lack these advantages. The structural complexity that makes these model systems valuable for studying development, behavior and disease also makes them impossible to freeze successfully.

It is a useful reminder that sometimes the simplest organisms are the most practically convenient – as long as your freezer is working properly.

Similar and related posts:

• The art and science of freezing bacteria
• Bacterial stocks and freezer failure Part I
• Bacterial stocks and freezer failure Part II
• Some experiences with freezer layouts

This story is not about one of those students.

The Setup

By March, our students have encountered Gram staining through multiple channels. They attend my Introduction to Microbiology lecture, where I explain both Gram and Acid Fast staining, highlighting certain similarities in the principles of the procedures. They then have a compulsory introductory lecture specifically for the practical, going through Gram staining step-by-step again. They receive a practical handbook, also compulsory reading, which details the entire staining procedure. And finally, they perform the Gram stain themselves during the autumn practical, working in pairs, one staining a Gram-negative bacillus, the other a Gram-positive coccus.

After Christmas, they sit an MCQ exam testing their understanding of these procedures in detail.

Come March, they face the practical assessment: a microscope, a slide, an academic and a demonstrator. The task is straightforward: set up the microscope, visualize the specimen and tell us what you are looking at. We provide support and prompting. The specimens I provide are all Gram-stained bacterial samples. In fact, the good students know that, if you have an unknown specimen, you always start with the lowest magnification, 40× for our teaching microscopes, and I always acknowledge that this is correct, but that for the sample I have given them they will not see anything at that magnification, and that they should switch at least to the 400× magnification to start. I am always surprised how little students deduce from their surroundings. I appreciate that they are nervous. However, I only teach microbiology topics and I only do the microbiology lab practical with them. The slide they are handed has, depending on what it is, a red or a purple sticker on the side so that it is easier for me to see whether they got the classification right. The prompt that they need at least the 400× magnification to see anything meaningful. Paying close attention these are valuable clues that students could pick up on. But so few do. The number of times students, when finally visualising the bacterial sample, told me: "This is lung tissue!"

The Assessment

One particular student arrived at my station very happy and quite chatty. I immediately liked chatting with him. Unfortunately, his knowledge of the Gram staining procedure was rather limited. There was some understanding – fragments, really – but extracting it required considerable work. Specific questions. Prompting. Gentle corrections when answers went astray.

He knew very little about what the dyes actually do mechanistically or how the procedure functions at a molecular level. But, with substantial guidance, he eventually arrived at the correct classification for his specimen: it was a Gram-negative bacillus.

Progress. And he seemed proud that he got there in the end.

We then moved to a different practical he had done: the blood smear. Students had prepared horse blood smears in an earlier practical, staining them with eosin and hematoxylin. The key indicators of understanding here are relatively straightforward: red blood cells lack nuclei and appear pink-red from eosin staining, while white blood cells display purple-blue hematoxylin-stained nuclei. This contrast makes distinguishing red from white blood cells simple, and nuclear morphology allows classification of white blood cells into their various types: neutrophils with multi-lobed nuclei, lymphocytes with large round nuclei, monocytes with kidney-shaped nuclei, and so on.

The student began describing his blood smear. He remembered preparing the horse blood slide. I probed his understanding of the staining procedure's purpose and mechanism.

Again, knowledge was limited. Very limited. But he was clearly trying, genuinely engaged with the material despite struggling to recall specifics.

The Breakthrough

When I asked about the colours of red and white blood cells, he paused. Thought hard. The silence stretched. I could see him working through it, searching his memory for the connection.

Then something clicked. His face lit up with recognition. A proud smile appeared – the unmistakable expression of someone who has just made what they believe to be a brilliant intellectual connection. With a tone of genuine achievement and complete conviction, he announced:

"Pink! Red blood cells were stained reddish-pink! So that means they are Gram-positive!"

Somehow I felt heart-broken for failing him. But, in part by being forced by the demonstrator, who struggled to comprehend what she just had witnessed, there really was nothing else I could do.

If you enjoyed this blog post, you might also enjoy the stories posted here, here and here.

Wishing all a Happy Easter 2026.

Liner AI offers support that appears to be geared towards academics and is supposed to help with research as well as writing. If writing is selected, Peer Review can be found in the list of agents on the left.

The service offers some free credits, so I decided to see what happened when artificial intelligence evaluated real scientific work. What came back was fascinating – and revealing about both the capabilities and fundamental limitations of AI in scientific peer review.

First impressions: almost too good

The AI review looked not bad on first glance.

It included:

• A structured meta-review summarizing strengths and weaknesses
• Five separate "reviewer" perspectives (Novelty, Rigor, Clarity, Impact, and Limitation)
• Specific recommendations with references to manuscript sections
• Balanced, measured tone throughout that we always see with AI these days
• Citations to relevant literature

This apparent quality is precisely what makes it interesting. And potentially dangerous.

The core problem: building a different paper

Once you read the AI review carefully, a pattern emerges. A large fraction of the recommendations fall into this category:

• You should measure mutation rates
• You should quantify DnaA protein levels
• You should perform transcriptomics
• You should test alternative antibiotics
• You should include proteomics
• You should determine MIC values
• You should test division mutants
• You should add time-resolved kinetics

The list goes on. And on. And ON.

Here is the fundamental issue: the AI assumes that if something is interesting, the paper must fully explain it mechanistically. That is not how real science publishing works – or at least not how it should work.

Our manuscript is a physiological observation combined with a methodological warning. It is cell biology and experimental methodology. It was never meant to be a complete mechanistic dissection, quite simply because we neither have the funds nor the manpower to develop it further than we currently have.

A human reviewer typically asks: Is the conclusion supported? Is the scope reasonable? Is the claim appropriate given the evidence?

The AI instead asks: What would the perfect paper on this topic look like?

Those are very different questions. The first evaluates what you have done. The second imagines what you could have done if you had infinite time, infinite funding and infinite graduate students.

The constraint blindness

The AI review repeatedly suggests experiments that are scientifically reasonable but operationally unrealistic:

• Perform single-cell transcriptomics
• Conduct proteomics on drug-treated populations
• Measure genome-wide mutation frequencies
• Screen other drugs systematically
• Include comprehensive dose-response curves
• Add viability assays at multiple timepoints

None of these suggestions are necessarily wrong in a wider sense. But together, they describe a project that would take many additional years, require substantially more funding – and likely never get finished.

The AI never asks: Is this needed to support the claim? Is this within reasonable scope? Is this realistic for one paper? Would this become a different project entirely?

This reflects a real problem in modern scientific publishing: the creep toward demanding, endless additional experiments before accepting work. AI reviews may unintentionally accelerate this trend because they optimize for completeness rather than feasibility.

A human reviewer with experience often thinks: "This would be nice, but it is not necessary for the points the authors are trying to make." The AI thinks: "If it is possible it should definitely be done."

When AI hallucinates criticism

More troubling, the AI sometimes criticized things that were already in the manuscript. This happens because AI review relies on pattern matching rather than genuine comprehension. It detects topics – replication, antibiotics, microscopy, statistics – and generates a generic checklist of expected issues. Some match real problems. Others are hallucinated relevance that sounds plausible but does not apply.

This makes the review look thorough while being partly disconnected from what is actually written.

Comparing to Human Reviewers

We have had this manuscript reviewed by human referees previously (an earlier version, admittedly). The contrast is instructive.

What human reviewers did:

Reviewer 1 made a specific technical disagreement. Right or wrong, this was an informed, contextual critique based on understanding how the drugs under investigation work. They questioned our experimental regime, asked for controls, and made specific procedural criticisms.

Reviewer 2 wanted mechanistic experiments, framing this as a "major concern" about understanding the molecular mechanism, a point we never intended to fully address – see funding, time and manpower constraints described above.

Both human reviewers also engaged in citation nitpicking, terminology arguments and extensive writing complaints (sentence structure, figure legends, terminology consistency).

What the AI did differently:

The AI provided more systematic coverage – every major experimental approach got evaluated against an apparent internal checklist. It identified some weaknesses we had possibly missed, such as unclear statistical reporting in some figures, insufficient imaging parameter details and other points.

However, here is the fly in the ointment: AI showed no contextual understanding. It did not grasp that this was a physiological observation paper, not a mechanistic study. It could not distinguish "this would strengthen the claim" from "this would be a different paper." It never made specific technical disagreements based on understanding bacterial physiology. It is something that I find more and more worrying these days.

Just as one example that always crops up in progression reviews for my PhD students. Let's assume we are measuring UV survival rates of very sensitive deletion mutants. Strains lacking the RecA recombinase, the main recombinase in E. coli, are exquisitely UV sensitive, several orders of magnitude below the wild type. Can I do a statistical evaluation of this difference? Sure! Do I actually have to do this? Well, my colleagues seem to think so, mainly because as part of their work they often get differences not particularly large, sometimes with error bars clearly overlapping. Here the question makes sense: are the differences observed likely to be significant? Statistics will provide a useful frame of reference to answer this question. But for the example of a strain that is consistently exquisitely sensitive to UV, with survival rates 1000-fold below wild type or more? It is telling that in the older papers that investigated survival rates for such mutants p values are very rarely shown. It is a waste of time. The effect is so significant that the answer is indeed obviuos. Asking for a statistical evaluation shows nothing but inexperience of the reviewer/AI.

But I know from experience that these days authors rarely can get away with this approach. So, I ran into the following almost comical situation. I calculated p values for many data points, all way below 0.001. When I ran the manuscript again, the AI review now complained, quite rightly, that I had used the wrong test for the data. So, I ran another set of tests on the same datasets. It was clear before I even started the tests that the results were highly significant. The wrong tests produced p values in the area of 10^–56, so I had only given "< 0.001" in the figures. So, I now ran test after test, without any impact on the work, with the exception of the statement which statistical test was actually used. But it made the AI review happier.

What the AI actually got right

To be fair, portions of the AI review were genuinely useful.

It correctly identified vulnerable points: mechanistic links that were speculative, sections with low cell numbers due to lysis, places where heterogeneity needed more discussion. It pushed for clearer -defined limitations, more precise wording and improved framing of contributions.

These are exactly the kinds of editorial feedback a competent reviewer provides. The AI is not producing nonsense only. It is producing a mixture of good editorial suggestions, generic checklist criticism and entirely unrealistic expansion requests.

The challenge is separating useful from impractical.

The Journal Scope Problem

Perhaps most tellingly, the AI review reads like evaluation for a Nature paper, not a specialized microbiology journal. It does not understand journal scope, impact level, or realistic expectations for different publication venues.

It evaluates every manuscript as if it should be complete, mechanistic, quantitative, clinically relevant, translational and future-proof. Real science publishing does not work that way. Different journals serve different purposes. Not every paper needs to be definitive.

But there is also a broader issue. AI tools meant to generate a review will generate a review. If you ask you will get an answer. Especially if you pay money for the service, the one thing that no one really wants to see is: "Great paper. Nothing to complain about. Good to go." It has to find points to correct, and pushing towards more is an easy way to achieve this.

Where this leaves us

My assessment: Technically impressive, editorially unrealistic, occasionally useful.

The AI review tool produces professional-looking output that would be genuinely helpful as a pre-submission checklist. "Have I addressed these standard concerns?" But it is not reliable as decision-making review because it lacks:

• Contextual understanding of scope and feasibility
• Ability to distinguish core claims from tangential extensions
• Recognition of resource and timeline constraints
• Domain-specific expertise for technical disagreements
• Understanding of journal-appropriate standards

Human reviewers have their own problems – inconsistency, bias, occasional incompetence, citation territorialism, writing style obsessions. But they can usually distinguish "this claim needs support" from "this would be interesting to know" and "this is feasible" from "this is a different project."

The AI cannot make those distinctions. Yet. It optimizes for an idealised version of every paper that may not exist in reality.

The practical takeaway

Should you use AI peer review tools? Perhaps, with clear-eyed understanding of what they provide.

Useful for:

• Identifying standard weaknesses you might have missed
• Checking statistical reporting completeness
• Flagging unclear sections
• Generating a pre-submission checklist
• Spotting missing methodological details

Not useful for:

• Deciding whether claims are supported
• Evaluating appropriate scope
• Understanding realistic resource constraints
• Making domain-specific technical judgments
• Replacing human editorial decisions

AI peer review is an interesting diagnostic tool. It is not a replacement for human expertise – and will not be until it can distinguish between "this would make a better paper" and "this would make a different paper."

In the meantime, we still need humans who understand that not every observation requires complete mechanistic dissection, not every interesting finding demands proteomics, and sometimes "we don't know why yet" is an acceptable answer if the observation itself is solid.

If you enjoyed this blog post, you might also enjoy this review.