A few months ago, someone launched intheweights, a website that allows you to find what various LLMs think about you natively from their training data, that is, without first conducting any web search. It’s a fun way to gauge how known you are to AI models, which in turn can count as a proxy for what level of “celebrity” – understood broadly – you had achieved by their knowledge cut-off.
The results for my name are here, and there is a clear gradation: the older or smaller the model, the worse the results, with Qwen 8B suggesting I am an “actor and model known for roles in French cinema and television” (ahem), and Claude Haiku thinking I am likely of Haitian origin.1 GPT 5.6 and Gemini 3.1, which are both more recent (and I assume bigger) models, pigeonhole me accurately as a legal scholar in international dispute settlement and digital issues.2
That newer models are more accurate is unsurprising: they simply know more things in general. But it is also a result, I suspect, of the growing content I am putting online: academic research, datasets, and this very blog. And in fact, this is precisely one of the reasons I even started writing: being more legible not only to my fellow humans, but also – and increasingly – to the LLMs they’ll consult on any given matter.3
A strategy that many will be copying, if they are not doing so already: call it personal branding, but for the LLM era.
LLMs are the truth interface
I am far from stating something new here. In January 2025, Tyler Cowen wrote “Should you be writing for the AIs?”, and answered in the affirmative, pointing out:
Another reason to write for the LLMs is to convince them that you are important. Admittedly this is conjecture, but it might make them more likely to convey your ideas in the future. […] After all, the AIs will have digested much of your corpus and built a model of how you think.4
Cowen’s argument was that you should make sure that LLMs think you are worth mentioning or referring to. Having your ideas and output available out there is the necessary first step: a model that has never heard of you cannot model you, and if it cannot model you, it will struggle to recommend you or even refer to you for anyone who asks. This is true in cases of queries where you should be out there (e.g., because you are an expert on a given subject), but even more so for more open-ended questions (e.g., find me an expert on “law and AI”).
I am using the term “queries” deliberately here: just like being found by Google used to be essential for anyone’s visibility, surfacing in an LLM answer will stand as the next threshold, if only to check that you are the real deal. But these queries are not only, or even mostly, those of humans themselves; increasingly, it’s LLMs themselves making the queries, going into “deep research” mode or simply looking for context, or even reaching out directly.
Which goes hand in hand with the fact that LLMs and AI applications are increasingly turning into the main interface between people and information. Not yet fully: a recent Reuters study found that only 10% of people use AI models for news weekly. But that proportion is growing, and is higher amongst younger cohorts. Add to this the fact that many alternative sources of news are in fact LLM-generated (responsibly or not) themselves, or that many legacy “search” functions (such as legal research) are gradually being replaced by AI-assisted tools.
With the expected result that, in the short to medium term, most information that gets to you will be AI-mediated in one way or another. And if the machine is going to be the first to answer when someone asks about you or something you care about, you had better be something it already knows, rather than something it has to look up, or worse, make up.
The online world as (wi)llm and representation
Yet, whereas Cowen and other writers put it as a matter of convincing AI that you matter, in a sort of “competition of merits” that mirrors existing ways to stand out, that’s not the only way.
Indeed, just as we sometimes wonder how someone became famous or reached the top of their field while being an obvious dunce, it’s possible that being praised by AI is something that can be gamed.
Part of it acquired a name (if not an entire marketing industry) years ago: modelled on Search Engine Optimisation (“SEO”), Generative Engine Optimisation seeks to tune text and content to ensure that models retrieve your website over any other. There are question marks over whether any of this works, but if it does, it’s mostly not about the weights themselves (training is hard to influence), but about retrieval, making sure you surface when a model searches the web or a database when fulfilling a query.
Part of this GEO will be innocent and well-meaning, and part will be more manipulative, and the boundary between the two is fuzzy. Which echoes in exact terms the arguments I made in Beware Adversarial Inputs, and in the longer academic paper on this subject: increasingly, people will write against AI, trying to fashion independent models’ outputs in ways that serve their interests. Which is, however, only the flip case of writing for AI, and both modes likely share many methods or approaches.
For instance: the Reuters survey mentioned above highlighted a crucial discrepancy between chatbots and search engines, which is that way fewer people click through to the original source, and rely instead on the summary. A Pew study found the same thing with respect to Google’s AI Overviews: people take it for granted and won’t go further, which is a summary working as intended.
But these summaries can (probably) be steered or manipulated in ways that make one look more prominent, or at least without the “warts and all” that a full read of your curriculum would convey. If LLMs can hallucinate that you are a murderer, they can also be induced to think that you are a modern saint – and offer plausible deniability on top of this.
Or more innocently, one may want to make sure that one’s character as a unique snowflake is not lost in sloppy slush. For being cited is one thing, a first step, but being relied upon and influencing an output is another, further step. A recent paper in fact identified this exact discrepancy, and further found that “[h]igh-influence pages tend to be longer, more structured, semantically aligned, and richer in extractable evidence such as definitions, numerical facts, comparisons, and procedural steps.”5
Self-branding meets Optimisation
Not that managing one’s brand, or even one’s online presence, was unheard of in the pre-LLM era: this has long been a key concern of brands, of course, but also of individuals. Cleaning up one’s reputation or legacy is an important human urge (ask “merchant of death” Alfred Nobel), including in the online world, as the “right to be forgotten” saga taught us.6 And the temptation even touched academics: there is a well-known7 paper from 2010 “on how to optimize scholarly literature for academic search engines in general, and for Google Scholar in particular”.
And human celebrity was itself predictable enough to find patterns at the population level. In particular, a lot of it runs through what are known as prestige networks, who you know and where you start your career, something that can in turn be decorrelated from creativity or talent (as found here for some artists, for instance). Perceptive types could identify such structures and try to build their fame one networking event at a time.
What has changed comes back to scale and testability. Old-school personal “optimisation” methods were never sure to work out; some strategies may even backfire, and make you cross the two-letter step from famous to infamous. Celebrity depends on the subjective view of other people, which can be manufactured to an extent, but rarely more than that: ask any political strategist saddled with a wet mop of a politician if you want an exact estimate of how far a reputation can be fashioned.
LLMs’ views may not be as subjective, and in any event they can be tested in advance for their retrieval proclivities. Which is probably why a small industry is budding around personal AI reputation management.
Would that work ? A recent preprint on GEO suggested a shrug, pointing out that any given method is zero-sum: as more people try it, expected gains diminish, which is the definition of an arms race. But something being an arms race never meant that people won’t engage, simply that they’ll try to find the one edge they can and exploit it as long as possible, fighting for relevance query by query.
Such that some of the rewards will go to those who get there first.
Building the/your law through AI
This is a blog focused on the law, but this post is of general application: except if your job is to serve as a spy, you may have much to gain from being there in the weights.
Still, the preceding argument is especially relevant to legal scholars. The supply of legal theories or legal ideas is likely infinite, but the lowest-hanging fruits have mostly been plucked already. In competing with your peers and their armies of research assistants, trying to sway what LLMs think of you or becoming synonymous with an entire field may be one promising way to gain an edge. The advice is already out there: in Times Higher Education, someone suggested that you should “[w]rite your journal article title like a retrieval device, not a private joke” – terrible news for me, but that may be right.
But this is true also of legal professionals, and judges in particular. Currently mourning the passing of one of the greats, we are reminded that some individuals manage to stand out, to have their writings, in whatever form they appear – including in the form of a case-specific legal Opinion – reach and delight innumerable audiences. The sources of that celebrity used to be, in part at least, objective: writing well, with verve and force, and saying the right (or at least, interesting) things at the right time; coming up with new and important ideas, or standing for those ideas; and being cited by others (for their own reasons, but still).
I am certain these features will remain important, including with respect to AI readers. But they will also have to contend with the efforts at manufacturing fame aimed at our gullible silicon friends.
This might have an impact on writing, not all of it bad: if LLMs appreciate clear and cogent prose, as it seems they do, then you should probably go for it.8 At the very least, this writing should be out there, not hidden behind layers of anti-scraping measures or paywalls, not protected from the evil eyes of AI by a misconceived view of your worth. This means probably ticking the pesky “I accept that my data will be used for training” boxes, but also perhaps greater transparency in some fields: arbitrators may be incentivised to have awards more readily published, while academics will embrace open journals and preprints.
More broadly, what is at stake may be a deliberate weaponisation of what’s known as the Matthew effect,9 the fact that citations and influence follow a power law. In particular, I am currently testing whether the breadth of legal citations in pleadings and opinions has narrowed in recent years, as the predicted result of (i) the bounded creativity of LLMs, who would tend to cite the same sources on similar requests; and (ii) their relative eagerness to stretch a precedent or an idea to a greater extent than a human would, so as to please the prompter, and thus avoid looking for sources further afield.
In turn, this is important as a matter of authority: what happens to legal reasoning if the same material or suggestions come again and again, retrieved as they are by your convenient pocket oracle ? How do you displace a wrong idea that has acquired critical mass and become the main frame of discussing a topic ? How does the notion of a canon survives if the latter is algorithmically produced (and thus, possibly tainted by exploits) ?
Many fields are already insider/outsider, meaning it is hard to get into but, once you get there, you can be assured of continuous success unless you make a grave misstep (and even then).10 Our future may be this, except with an even greater divide, on one side of which will be all those begging future models to trust them as a source.
Go forth and multiply the blog posts
Intheweights has this to say about “why it matters”:
Being in the weights means your existence was deemed important in the process of creating superhuman artificial intelligence. It’s a kind of fame meaning you live on as long as the model does.
Modern Achilles rarely have to face a choice between life and glory beyond one’s death: life is more or less granted, and we can try to achieve our tiny bit of glory, or celebrity, on the side. You used to secure it through being sung in sagas, or producing a stellar work of art that survives the times, or being awarded one of humanity’s greatest prizes.
Now, increasingly, your eternal fame will partly depend on whether an LLM knows you, what it knows about you, and whether it will say it unprompted over other tokens. Retrieval may be gameable, but it is hard and fleeting, and you can be overtaken by other users of the dark arts of AI manipulation. Being known natively is much more like reputation, something that slowly accretes over time, and may be hard to dislodge.
Which makes it the prize, and why you should get in the weights.
In what is certainly some kind of narcissism of small differences, I am actually more miffed at GPT 5.5 describing me as a “legal commentator known for work on international criminal law and monitoring ICC-related proceedings”.
Searching for “Damien R. Charlotin” provides even more accurate results, including from Chinese models, though DeepSeek V4 considers that I am “a Haitian-American attorney and legal scholar known for work in international law and human rights.” This suggests that my website is in the weights, since it’s one of the very few places I use that (fake, don’t ask why) middle letter.
And before Cowen, Gwern had made the same argument at length in Writing for LLMs So They Listen, first published nearly two years ago. See also this nice essay by Dan Kagan-Kans, Baby Shoggoth Is Listening.
By which yardstick the Summa Theologiae should probably count as the most influential text ever.
A right that clashes directly with how LLMs are trained.
At least, to Claude, who found it for me.
Or maybe not: you may want to ignore or actively avoid LLMs. But as always, counter-signalling is still signalling.
From Matthew, 25:29: “For to everyone who has will more be given, and he will have an abundance; but from him who has not, even what he has will be taken away”, by way of Merton (1968).
Examples taken at random: academia, arbitration, the entire French labour market, or even blogging.