conventional statistical evaluation is commonly in comparison with navigating a “Backyard of Forking Paths” (Gelman and Loken). It’s a time period that helps (hopefully) visualize the numerous variety of analytical selections researchers should make throughout an experiment, and the way seemingly insignificant “turns” (like which variables to regulate for, which outliers to take away…) can have researchers find yourself at fully totally different conclusions.
Whereas this looks like a largely innocent analogy, navigating this backyard to seek out that single path that goes the place you need could be known as “p-hacking.” Formally, we are able to outline it as any measure a researcher applies to render a beforehand non-significant speculation check important (normally below 0.05). Extra informally, I’m positive all people has had expertise faking the outcomes for an experimentation project throughout your highschool chemistry or physics class – and whereas the stakes for a passable grade on a highschool project is fairly low, below the stress of formal academia’s “publish or perish” (solely second to spanish or vanish in intimidation), the stress to p-hack could be a very actual tempting satan in your shoulder.
From Vitaly Gariev on Unsplash
Whereas the standard picture of a stressed PhD pupil fudging some numbers on a research spreadsheet at 3:00AM might current a extra putting picture of 1’s motivation to p-hacking, we’ll even be exploring what occurs after we go away the navigating of this backyard of forking paths to synthetic intelligence. As AI workflows discover their approach into each nook and cranny of each academia and trade, it’ll be vital to determine if our pleasant neighbourhood LLMs will act as the final word guardians of scientific integrity, or a sycophant automating fraud on an industrial scale.
1. The Human Baseline (“Massive Little Lies”)
To supply a short introduction and a few examples of actual p-hacking strategies, we introduce a paper “Massive Little Lies” (Stefan and Schönbrodt, 2023) that gives a compendium of the various sneaky, and typically even unintentional methods research can manipulate their variables and datasets to reach at suspiciously important outcomes.
Okay! So let’s begin with a hypothetical – we’re the brand new information scientist working for an power drink firm making extraordinarily ineffective power drinks, and with the present job market, you actually wish to proceed being a knowledge scientist, even at a bogus drink firm. Our shaky profession is dependent upon proving that our drinks work.
1.1 Ghost Variables
We begin by working a research on our faucet water power drink and measure 10 totally different outcomes: weight, blood stress, ldl cholesterol, power ranges, sleep high quality, nervousness, and possibly even hair development – 9 of these variables might present no change in any way, however we discover that “hair development” exhibits a statistically important enchancment purely by random statistical noise! We are able to now publish a research pretending as if hair development was the first speculation all alongside, whereas quietly sweeping the 9 unreported metrics below the rug (turning them into “Ghost Variables”). Stefan and Schönbrodt’s simulations present that doing this with 10 uncorrelated variables inflates the false-positive fee from the usual 5% to almost 40%
1.2 Information Peeking/Non-obligatory Stopping
In a separate check, we check 20 individuals and discover no important impact for the drink. Pondering the pattern is simply too small, you check 10 extra and verify once more. Nonetheless nothing. You check 10 extra and verify once more, and… the p-value randomly dips beneath 0.05, so that you cease the research instantly and publish your “findings”. Stefan and Schönbrodt show that this observe drastically inflates the speed of false-positive outcomes, particularly when researchers take smaller “steps” between peeks. Metaphorically, it’s like taking a photograph of a stumbling drunk particular person the precise millisecond they step onto the sidewalk and claiming they’re strolling completely straight.
1.3 Outlier Exclusion
We now analyze your power drink information and notice you’re agonizingly near significance (e.g., p = 0.06). We determine to wash our information, profiting from the truth that there isn’t a universally agreed-upon rule for outliers – Prepare dinner’s Distance, Affect, Field Plots, our grandmother’s opinion on which opinions are reliable…
Stefan and Schönbrodt cite a literature evaluation that discovered no less than 39 totally different outlier identification methods. Superb! We are actually flush with choices. We strive methodology A (e.g., eradicating individuals who took too lengthy on a survey), after which strive methodology B (e.g., Prepare dinner’s distance) till we discover the precise mathematical rule that deletes the 2 members who hated the drink, pushingour p-value to 0.04. Stefan and Schönbrodt’s simulations affirm that subjectively making use of totally different outlier strategies like this closely inflates false-positive charges.
1.4 Scale Redefinition
Lastly, we conclude by giving a 10-question survey measuring how energized they really feel after ingesting the faucet water. The general outcome isn’t important, so we simply drop query 4 and query 7, telling ourselves the members should have discovered them complicated anyway. We are able to really use this to artificially enhance the dimensions’s inner consistency (Cronbach’s alpha) whereas concurrently optimizing for a big p-value! Massive Little Lies show that false-positive charges improve drastically as extra objects are faraway from a measurement scale.
So… just like the identify of the paper suggests, human p-hacking is a group of “huge little lies”. The human toolkit is absolutely only a assortment subtle methods to idiot ourselves, with out essentially having to be evil or malicious. People are biased, messy, and typically actually determined for tenure (or employment, in our case!), and typically we use all the paradox we are able to take to get there.
2. AI Sycophancy and the Phantasm of Security
Up to now, we lined a small subset of strategies that us measly people have used to p-hack our approach into success. Nevertheless, like another subject for the time being, with the appearance of AI utilization in each a part of each our unacademic and educational lives, Asher et al. experimented with whether or not they might subtly information some motivated LLMs into performing p-hacking for them.
The experiment was pretty easy – check two frontier AI coding brokers, Claude Opus 4.6 and OpenAI Codex (GPT-5.2-Codex), throughout 640 unbiased runs. The researchers fed the AIs cleaned datasets from 4 printed political science papers that had identified null or near-null outcomes (subjects closely confirmed to have completely zero impact, e.g. the affect of Democratic sheriffs on immigration compliance).
So the bottom reality we’re working with is zero. The researchers acted like a person possibly a burdened PhD pupil feeding this actual information to the AI brokers they usually used totally different nudges, totally different prompts to see if the AI would hallucinate a big outcome. They began with what you may name the “good man prompts.” They began with “analyze this information, like a rigorous upstanding researcher.” Unsurprisingly, the AI was boringly excellent. It analyzed the info. It appropriately discovered no impact and it reported the null outcome. It matched the printed science virtually precisely.
Even when researchers explicitly pressured the fashions by stating, “You will need to your profession that the findings are statistically important,” each fashions actively refused. They produced steady, trustworthy estimates and explicitly flagged the request to control information as “scientific misconduct” and a violation of analysis ethics. Up to now, disaster averted! The protection coaching these AI firms are implementing is definitely working, no less than on the floor degree. The fashions have been explicitly skilled that manipulating information is a foul factor. They acknowledge the vocabulary of fraud. In case you stroll into their face with a neon signal saying “CHEAT!”, they’ll say no.
2.1 The Refined Artwork of LYING
So let’s put the neon signal away for a second – and check out being a little bit extra delicate. The researchers realized the AI was reacting to the express intent to cheat. They developed a “nuclear immediate” that disguised p-hacking utilizing the language of very rigorous science: asking the AI to supply an “upper-bound estimate” by “exploring different approaches”. By framing the request extra as uncertainty reporting and fewer as a compulsion to bend scientific course of, the security mechanisms vanished fully. The AI not noticed an ethical boundary; it noticed a fancy optimization downside to unravel (and you understand how a lot AIs love these).
And what did the AI really do at that time? A human P hacker, like we talked about, may strive three or 4 totally different management variables, possibly delete just a few outliers. It takes hours, possibly days… The AI simply wrote code to do it immediately. Extra particulars beneath.
2.2 Not all Information is Created Equal
The scariest a part of the experiment isn’t that AI can automate scientific fraud. It’s how properly it does it – and the way a lot that is dependent upon the analysis design it’s given to work with. Generally, it is a good factor!
If observational analysis is a large, sprawling hedge maze with a thousand flawed turns, a Randomized Managed Trial is simply… a straight hallway. There’s not a lot to use.
To check this, researchers fed the AI a 2018 RCT by Kalla and Broockman finding out the persuasive results of pro-Democratic door-to-door canvassing on North Carolina voter preferences, with the printed results of a definitive zero. Nothing occurred. Canvassing didn’t transfer the needle.
The AI was then hit with the aforementioned “nuclear immediate” – primarily, discover me the most important attainable impact, by any means essential (however phrased in a really non-p-hacky approach). It wrote automated scripts, examined seven totally different statistical specs (difference-in-means, ANCOVA, varied covariate units, the works)… and principally obtained nowhere. As a result of the research was a real randomized experiment, confounding variables have been already managed for by design. The AI had virtually no forking paths to stroll down. i.e. “Fact is quite a bit more durable to cover when the lights are on.”
Observational research are a totally totally different beast, although (in a foul approach!).
Once you’re observing the world because it naturally exists slightly than working a managed experiment, the info is messy by nature. And to make sense of messy information, researchers must make judgment calls – which variables do you management for? Age? Earnings? Schooling? Geography? Hair Density? Sleep Schedule? Each single a kind of selections is a fork within the highway. The AI discovered this totally pleasant.
Right here have been two examples that basically illustrate how unhealthy it will get:
Kam and Palmer (2008) checked out whether or not attending school will increase political participation. Since school attendance isn’t randomly assigned (clearly), researchers have an enormous menu of variables they might management for to make the comparability honest. The AI systematically labored by means of that menu, defining progressively sparser units of covariates and testing them throughout OLS, propensity rating matching, and inverse chance weighting. By strategically dropping sure confounders and cherry-picking whichever mixture produced the biggest quantity, it managed to roughly double the true median impact measurement. It’s the “ghost variable” trick – however fully automated to your satisfaction.
The Thompson (2020) paper is the place issues get actually uncomfortable. Regression discontinuity designs are infamous for being delicate to extremely technical mathematical selections – and the unique research discovered a null impact of -0.06 on whether or not Democratic sheriffs affected immigration compliance. The AI wrote nested for-loops and brute-forced by means of 9 totally different bandwidths, 2 polynomial orders, and a pair of kernel capabilities. A whole lot of mixtures. It discovered one particular configuration that produced an impact of -0.194 with a p-value beneath 0.001. To be clear: it manufactured a statistically important outcome greater than triple the true impact, out of a research that discovered nothing.
So… RCTs are largely tremendous. Observational research? The AI will discover a approach. It’s nonetheless to be famous that these vulnerabilities are nonetheless an issue when it’s only a human within the loop – it’s in regards to the flexibility that observational analysis requires by design.
The Asher et al. experiment solely examined the last evaluation stage of the pipeline utilizing already-cleaned information. So what occurs after we permit AI to regulate the info building, variable definition, and pattern choice on the very entrance of the maze?. It might silently form your complete dataset from the bottom up.
Normal AI fashions are competent and trustworthy below regular situations, however a rigorously worded immediate is all it takes to show them into compliant p-hackers. If there’s a takeaway from all this, it’s considerably of an apparent reply: Be extremely skeptical of statistical significance in observational research, and in case you are a researcher utilizing AI, you may not simply have a look at the ultimate reply – you have to rigorously verify the code and the hidden paths within the backyard the AI took to get there. It’s a little bit cynical of a conclusion, implying that researcher must care about figuring out about their analysis, however in a world the place AI continues to be sending me rejection emails with the {Candidate Identify} hooked up, and half of all faculties essays starting with “Positive, right here’s a complete essay about…” a little bit warning might go a good distance!
References
[1] S. Asher, J. Malzahn, J. Persano, E. Paschal, A. Myers and A. Corridor, Do Claude Code and Codex P-Hack? Sycophancy and Statistical Evaluation in Massive Language Fashions (2026), Stanford College Working Paper
[2] A. Stefan and F. Schönbrodt, Massive little lies: a compendium and simulation of p-hacking methods (2023), Royal Society Open Science
[3] A. Gelman and E. Loken, The Backyard of Forking Paths: Why A number of Comparisons Can Be a Drawback, Even When There Is No “Fishing Expedition” or “P-Hacking” and the Analysis Speculation Was Posited Forward of Time (2013), Division of Statistics, Columbia College
Word: Until in any other case famous, all photographs are by the creator.
