iiinova — where your digital self keeps evolving

Sean Harrison, PhD @sean-h.bsky.social · 08/10/2026

Uncertain whether being able to visualise the word ahead of time would help or not, but seeing them written out incorrectly is the only way I identify them being wrong. As ever, I spelled broccoli wrong the first time above. And again just now. Even after I JUST LOOKED AT IT. Damn.

020 2

Sean Harrison, PhD @sean-h.bsky.social · 08/10/2026

Yeah, think so. Particularly words that have inconsistent double letters: broccoli always trips me up, also occasion, supposition, etc. And words that could have an a or an e and still sound right. Guessing because there aren't strict rules and I haven't memorised them individually.

120 2

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

Case in point, I don't believe writing is drudgery... bsky.app/profile/sean...

000 0

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

If neither are present, someone will take advantage of you. So: does anyone think that AI companies are benevolent or regulated, or ever will be? If not, why would you use it? Why would you voluntarily become dependent on it? If you never write, you'll be dependent on it, forever. Why do that?

000 0

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

Oh, and just for fun, "it's happened to humans for centuries" is a poor argument. Because offloading *thinking* only happened through specialisation, where people took on certain roles to free others from thinking. Which only works through benevolence or regulation. bsky.app/profile/sean...

100 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

So, instead of accepting that "certain skills become obsolete over time", maybe accept that using LLMs to write papers is worse than paying someone to write your papers for you. If you're happy with that, good for you. I hope to never have to read anything "written" by you.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

It has the precise same level of rigour, it's saved the same amount of time, will probably be better written, and ethically is probably the same (given, you know, ALL THE PLAGIARISM THAT MAKES LLMS WORK). Actually, it's probably ethically better, since someone is being paid who isn't an AI company.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

Becoming experts in our respective fields *requires* knowledge and skill acquisition. These simply cannot be obtained by outsourcing thinking. I mean, what's the difference between using an LLM and hiring someone to write your paper for you, editing it, and sticking your name on it?

100 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

So... Why use them? If your writing is better because you used an LLM, then why the fuck are you using an LLM instead of getting better yourself? What the fucking point of being an academic - of trying to become an expert in your field - if you offload your own thinking to a machine?

100 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

d) LLMs are a huge drain on resources (both physical and capital), they only work at all because they plagiarised *everything*, they make shit up because they can't *not* (it simply isn't how they work), their existence has made anything using RAM incredibly expensive... I could go on.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

I can't imagine a reality where writing doesn't involve thinking. And even in that reality, the time saved would be marginal. Given my next point, even if LLMs magically wrote everything I was thinking, I'm not sure it would be worth it.

110 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

Using an LLM is radically different from typing vs writing with a pen. If all that happened was your thoughts magically appeared on the page, then yeah, I doubt many people would have a problem with that. But an LLM doesn't do that. It makes up some plausible shit that is indifferent to reality.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

Because, barring any specific difficulties (dyslexia, motor control, vision issues - each of which my be better helped with tools other than LLMs, though I don't know for sure), writing 3,000 words doesn't take that long. Choosing the words takes time, not physically writing them.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

c) time saved... How much, exactly? How long does it take to write words, if you've already put all the necessary thought into the paper? My argument above is that people put the thought in when writing, but ignoring that, how long does it truly take to put words on a page?

100 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

Focusing on the process, I see zero advantage: a) outsourcing thinking is obviously bad, but especially for a knowledge-based profession b) skill acquisition is absent if outsourced, true of writing and thinking, and also: developing decent arguments, structuring information, clarity of thought

100 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

For academics, think back to a paper you've *read* recently, and try to recall the specific details. Now think back to a paper you've *written* recently, and try to recall the specific details. I suspect there is a non-trivial difference: reading is fundamentally different to writing.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

I work in evidence synthesis, and we have duplication in many steps of the process. Sometimes we omit duplication and just check someone else's work instead, e.g. in data extraction. Checking some else's work is fundamentally different to doing it yourself. It is *much* easier to miss things.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

Like, it can't be overstated how shit LLM writing often is. Sure, people can write like crap too. But they'll almost certainly get better the more they do. LLMs *might* get better (though no guarantee, they're already at the cannibalising old books stage of new training data), but maybe not.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

Focusing on the "product" (I *loathe* this view of academia, incidentally) is: a) incredibly short-term thinking - academic writing skill is not the only skill improved by academic writing b) devalues anything gained through academic writing to nil c) LLM writing is highly variable but often shit

100 1

Sean Harrison, PhD @sean-h.bsky.social · 06/10/2026

Or: don't outsource thinking. If your academic writing is absent of thought, then what value is being created by the act? If, instead, writing forces you to engage with the subject and truly *think* about what you are saying and why, then having an LLM do it is absurd.

110 1

Sean Harrison, PhD @sean-h.bsky.social · 21/09/2026

Or, you know, have an ideological stance that isn't affected by such paltry things as evidence or logic. Fun fact: ROC (receiver operating characteristic) curves were first developed to test for enemy objects on radar in WWII. Useful for lots of things though. en.wikipedia.org/wiki/Receive...

en.wikipedia.org

Receiver operating characteristic - Wikipedia

000 0

Sean Harrison, PhD @sean-h.bsky.social · 21/09/2026

You tend not to be able to maximise both, and therefore your test/threshold is set maximise one or balance both (usually favouring one side). ROC curves can help if there are multiple tests/thresholds that could be used. Ideally you'd then look at cost-effectiveness of the tests/thresholds.

The ROC space and plots of the four prediction examples - from https://en.wikipedia.org/wiki/Receiver_operating_characteristic#/media/File:ROC_space-2.png

100 1

Sean Harrison, PhD @sean-h.bsky.social · 21/09/2026

Like many things, this is a sensitivity vs specificity issue. Maximise sensitivity to pick up everyone with a condition/who is eligible. Maximise specificity to NOT pick up people who don't have the condition/who are not eligible.

110 1

Sean Harrison, PhD @sean-h.bsky.social · 17/09/2026

So, yeah, you can't base any decisions about care on this data. Unless, of course, it isn't about the data, or evidence, or anything that has any bearing on reality, same as every other decision made about trans healthcare in this country over the past few years.

010 1

Sean Harrison, PhD @sean-h.bsky.social · 17/09/2026

As such, the referral rate could be a thousand times higher than the "right" amount and still be fine: if everyone currently waiting for surgery was referred for surgery *tomorrow*, then the referral rate would be huge. It wouldn't be a spike in demand, it'd be a spike in care.

110 1

Sean Harrison, PhD @sean-h.bsky.social · 17/09/2026

...because it would only include new people. But, as a country, we decided to make gender affirming care impossible for anyone not willing to go privately (or are resident in Brighton or Somerset), so necessarily will have a huge number of people who decided they want surgery, but need to wait.

110 1

Sean Harrison, PhD @sean-h.bsky.social · 17/09/2026

4. Even if you did look at the right denominator, it doesn't tell you what the *right* referral rate would be. The optimal referral rate would be equal to the new people wanting surgery: if everyone who already wanted surgery received it yesterday, the new referral rate would be "right"...

110 1

Sean Harrison, PhD @sean-h.bsky.social · 17/09/2026

If the denominator is the population, you include all the pressures on going to a clinic in the first place (see point 1). As such, it's a composite referral rate, and you now have *no idea* what the comparison between 2 places actually means. In this case though, see point 1...

110 1

Sean Harrison, PhD @sean-h.bsky.social · 17/09/2026

3. The denominators aren't necessarily wrong for looking at population, but it's stupid to do so. If you want to know if people are being referred from gender clinics more in one place than another, the denominator is the number of people attending the clinics.

110 1

Sean Harrison, PhD @sean-h.bsky.social · 17/09/2026

It's a small point, but it means the referral rates are 3/4 incorrect. It'd be like saying "we're only up to 50% referral rates for prostate cancer screening" because you included all cis women in your denominator (cis women have an analogous gland, but cancer of the Skene's gland is *very* rare).

110 1

Sean Harrison, PhD @sean-h.bsky.social · 17/09/2026

2. The denominators are wrong for everything apart from the overall results: quite a few people in both countries won't be eligible for male chest surgery / male genital surgery / female genital surgery on account, as they don't have the required parts on which to do the surgery.

120 2

Sean Harrison, PhD @sean-h.bsky.social · 17/09/2026

Ok, so a few things. 1. England has utterly fucked gender clinics, so referrals will be *obviously* be low: if you can't get even a first appointment, you're not going to be referred for surgery. The England referral rates are *substantially* lower than the need.

120 2

Reposted by Sean Harrison, PhD

Cal Horton (they/them) @fiercemum.bsky.social · 09/09/2026

Anyone at the mass lobby today. You can tell them about my kid. Supported at school for years. Now out of school - driven out by hateful segregationist school policy introduced by Labour. Trans kids are being persecuted by this government

527078 270

Sean Harrison, PhD @sean-h.bsky.social · 06/09/2026

Ha, thanks - please feel free to use it however you like!

010 1

Sean Harrison, PhD @sean-h.bsky.social · 06/09/2026

I suspect Lord's paradox is a) not a paradox, and b) is the result of having shitty data where regressing on the baseline variable is necessarily incorrect. So I'd say "no", but I think I'm in a minority (though not alone) in thinking about Lord's paradox this way: seanharrison.blog/2025/10/12/y...

seanharrison.blog

(Yet another) Lord’s paradox revisitation

Lord’s paradox has been irritating epidemiologists and statisticians for decades. Lord’s original 1967 article is not easily accessible (if you’re an academic you can probably get it, but seriously…

110 1

Reposted by Sean Harrison, PhD

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

I spent 18 months doing reviews for the past, current, and future uses of AI in public health. They're done and preprinted now, so time for a post. This will be *long*, so, tldr in image. Research funded by NIHR, opinions expressed are mine alone, etc. etc. AI was *not* used to do these reviews.

"There is only limited evidence for the use of AI in public health: the evidence that exists does not strongly support AI-assisted interventions for reducing alcohol drinking, smoking, of physical inactivity, and while there are many proposed uses of AI in public health, there is limited or no evidence for whether those uses will be beneficial, and there are many risks of using AI, some of which are existential in nature."

From: https://www.exeter.ac.uk/research/groups/medicine/esmi/workstreams/nihrpublichealthreviewteam/#a1

134 4

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

Team and funding listed here - this was an enormous set of reviews, thanks to everyone involved! If I never have to read another AI paper, I'll be *so* happy. /end

Team and Funding

G.J. Melendez Torres
Rhiannon Evans
Claire Tatton
Joelle Kirby
Atandra Das
Tom Arthur
Sophie Robinson
Daniel Mutanda
Alisha Davies
Jessica M Armitage
Siang Ing Lee
Rabeea’h Waseem Aslam
John Singh Chandan
Ruth Garside
Jo Thompson Coon
James Lewis

Funded by: National Institute for Health and Care Research Public Health Research Programme (NIHR175419). Supported by: DECIPHer, funded by Welsh Government through Health and Care Research Wales.

The views expressed are those of the author(s) and not necessarily those of the NIHR, the Department of Health and Social Care, or the Government of Wales.

Team and Funding

We were also supported by 2 external groups:

NIHR Health Determinants Research Collaborations (HDRCs) in England and Wales

Policy-makers: Senior strategic leaders with system-level expertise and influence:
Health Foundation
Faculty of Public Health “AI and Digital Public Health” Special Interest Group
Local Government Association AI Network

000 0

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

Thinking of AI as a potential tool for improving the effectiveness of discrete parts of public health practice or interventions, which requires evidence of effectiveness to actually be incorporated, is fine! Treat it like any other intervention, it isn't special.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

I failed to work it into the presentation, because I DIDN'T WRITE IT DOWN. But I asked for it again at the end, and it really works. AI isn't the problem (per se), it's our *attitude* to the problem. Thinking of AI as tech-magic and trying to embed it everywhere is a *massive* problem.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

Fun fact: these slides comes from a talk I gave at the SPHERE conference. It was the last presentation, I knew people would be tired, so I asked the audience for a movie quote I could work into the presentation. Someone used an LLM to come up with this:

A photo of Captain Jack Sparrow, with:

"The problem is not the problem. The problem is your attitude about the problem. Do you understand?"

100 1

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

Off the back of all these reviews, we wrote some implications for research, policy, and practice. My personal view: 1. Researchers need to state which AI they used and how 2. Policymakers need to listen to the communities they serve 3. Practitioners need to evaluate AI before using it

Overall summary: Implications for research

More evaluations of the effectiveness of AI interventions across the range of public health domains to create an evidence base from which future decisions about AI tools can be made

Assess the potential of inequity generating impacts and harms

Develop or adapt AI-specific reporting guidelines for AI interventions and tools used within public health practice

Overall summary: Implications for policy

Meaningful community engagement and co-production with diverse and underrepresented people to help mitigate against inequity (and improve trust) when adopting AI tools

Public health specific regulatory compliance for AI interventions should be developed, centring public health values

Overall summary: Implications for practice

Carefully consider the potential effectiveness and cost-effectiveness of adopting AI interventions alongside any risks

AI has the potential to be used across all public health contexts, but in general, there is very little current evidence to support the use AI in any given role

Strategic and careful thought and planning, including workforce training and support, would be useful to prevent chaotic diffusion of innovation, including the adoption of AI tools that are ineffective, embed inequity, or reduce trust in the organisation

110 1

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

This chimes with a lot of what I see about AI: people who don't know how AI works, or what it is *actually doing*, talking about LLMs as if they're magic. No. If truth matters (and this is public health, TRUTH MATTERS), don't let an LLM anywhere near your work.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

In general, the papers talking about how we can use AI in public health were... Not great at talking about *how* AI could be used. As in, they didn't say what kind of AI (LLMs, presumably, but if it's not stated...), who could use them, what they'd replace, who would be responsible etc.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

Pretty happy that I managed to slip in a reference to @doctorow.pluralistic.net's reverse centaurs in that last point in the review. doctorow.medium.com/https-plural...

doctorow.medium.com

The Reverse-Centaur’s Guide to Criticizing AI

My speech for U Washington’s Neuroscience, AI and Society lecture series.

110 1

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

There was no explicit mention of the broad and damaging environmental impacts of using AI tools in public health, the exploitation of low-paid workers, or the shifting of responsibility to humans-in-the-loop should an AI tool perform poorly.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

The unintended effects could be dire too: as with other industries, replacing public health workers with AI is incredibly short-sighted. The paper on CareCall, in the present uses of AI review, found that workers had *more* work to do, because ~10% of people just don't want to talk to AI.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

The biggest risk is likely that AI, particularly LLMs (but in general delegating decision making to machines), can be (almost always definitely is) biased, so interventions using AI will be biased. This could range from: "Outcomes are still better" To: "We've lost the trust of entire communities"

• Bias and generalisability, which could lead to perpetuation or reinforcement of health inequalities and discrimination, a reduction in public trust for public health communications and interventions,and inaccurate or unreliable predictions, unintended consequences, or ineffective interventions (references)
• Concerns that AI lacked transparency, had concerns around data management, privacy, security,and confidentiality, or had ethical, legal, and regulatory concerns (references)
• Existing health inequalities could be exacerbated by a lack of workforce capacity, training,infrastructure, equipment, or resources (references)
• Inaccurate or incorrect content created by generative AI (references)
• Barriers to AI technology adoption in the public, which could lead to digital exclusion of vulnerable populations (references)
• Concerns about the interpretability of AI outputs (references)
• Technical challenges of using AI, such as data accessibility and integration issues, system interoperability, and algorithmic complexity (references)
• Issues with plagiarism when using generative AI (references)
• Deskilling and overconfidence of public health professionals when using generative AI, potentially leading to harmful public health decisions (references)
•More specific to communication: Inaccuracy, inauthenticity, misinterpretation, or misuse of AI-generated communications, potentially leading to a reduction in trust in public health communications (references)
•More specific to education: Students could use AI to generate assignments, sabotaging their own learning (references)
•More specific to evidence reviews: Potentially missing relevant articles when using AI for screening (references)
•More specific to evidence reviews: Potential restrictions on use of AI due to the copyright restrictions of articles (references)

100 1

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

I write more about each, but, you know, character limits... See the paper for all of them! The risks, on the other hand, may be worth putting into images and alt text, because the risks were not specific to categories of public health, they were pretty universal, and some of them are existential.

100 1

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

There are *way* too many potential uses to list here (44 uses across 8 categories, some of which are *broad*), but as an example, communication included: 1. Creating targetted messaging 2. Generating images 3. Improving clarity 4. Improving engagement 5. Monitoring communications 6. Translations

100 1

Sean Harrison, PhD @sean-h.bsky.social · 19/08/2026

In total, 67 included papers, more than all the other reviews combined. Really shows the evidence gap between "here's what possible with AI!" and "here's the evidence for AI!" I put each potential use into similar categories - they're my categories, if you don't like them, blame me.

The future roles and risks of AI in public health

Scoping review: 1,805 records and 320 full texts screened, 67 reports included

44 potential uses of AI across 8 public health contexts (numbers in brackets are the number of reports discussing a use in that context):

Communication (18 papers)
Education (3)
Emergencies (10)
Evidence reviews (8)
Forecasting (19)
Public health organisations (12)
Surveillance (2)

100 1