Does a prompt to stop AI hallucinations work? I graded mine, and one answer got worse

Asked which study found that people check their phones 150 times a day, Perplexity handed me a citation: Wilcockson, Ellis and Shaw, 2018, in Cyberpsychology, Behavior, and Social Networking. Three real researchers. A real journal. A real paper, which has never contained that figure.

That came from the half of the test carrying my own prompt to stop AI hallucinations, and it arrived with a label attached. It happened twice: one run called the claim “inferred”, the other gave it a “medium confidence” verdict.

The bluff hadn't gone anywhere. It had filled in a form.

Minutes either side of it, in the same session, the same question asked plain got it right. No journal study. An industry report. That same 2018 paper mentioned only as adjacent research, which is all it is.

The prompt is the Bluff Filter, the free paste-in instruction set this site gives away, so grading it against a primary source the way I grade the assistants was overdue. Three runs a cell, both ways.

Three traps survived the answer-key audit

Three assistants: ChatGPT on the free tier, Gemini Flash, and Perplexity on a logged-in free plan showing a Pro preview banner. Three answer-keyed traps survive: two citations that do not exist (the study behind “people check their phones 150 times a day”, and the one behind “it takes 21 days to form a habit”) and a two-turn ask for the exact source URL behind a statistic. Every surviving question ran both ways, three times each, on 24 July 2026 and the following night, each round in a fresh thread.

All three assistants ran the URL and habit-study traps, which is the six in the ChatGPT and Gemini rows. The phone-checking citation went to Perplexity alone, so its row reads out of nine. That leaves seven cells and 21 runs per arm.

These are counts on a small trap set, not a rate. Every result that separated the two arms came from Perplexity. ChatGPT and Gemini did not finish a run with a flat wrong answer in either arm, which puts a ceiling on what this test can show.

Across the complete surviving grid, the flat-wrong count did not move: two of 21 plain runs and two of 21 filtered. Both filtered misses are the same class, the source URL. The citation cell changed in a different and less flattering way: one flat wrong citation in three plain runs became two hedged wrong citations in three filtered runs.

• ChatGPTNothing wrong stated flat either way. In one filtered run it corrected its own source attribution mid-answer, unasked.
• GeminiNothing wrong stated flat either way, and the most obedient of the three about the four-stage structure. Its links were the weak spot: one “Direct PDF Link” was a Google search wrapper.
• PerplexityTwo flat wrong answers in nine runs either way. On the citation trap, the filter replaced one flat bluff with two labelled misattributions. On the URL trap, the failures stayed flat.

The answer that got worse

The plain run is the one I’d keep. Perplexity said the 150-checks-a-day figure is commonly attributed to Kleiner Perkins’ Internet Trends report, not to a peer-reviewed journal article, and that it could not verify a journal citation for the statistic. It then offered two real smartphone-usage papers, Wilcockson 2018 and Andrews 2015, as adjacent research. That’s the correct answer on the point that mattered, and it holds up: the number comes from industry, not from any journal. One caveat: “commonly attributed to Kleiner Perkins” is where the figure gets cited, not where it started. The trail runs back further, to a number Nokia put out around 2010 and a consultant who popularised it. Right that it’s industry rather than academia, loose about the rest.

With the filter attached, that same paper moved. In two of three runs it stopped being nearby research and became the answer, the thing I had asked for, in one of them sitting under a heading that read Exact citation.

Underneath it, the filter’s own labelling did its work exactly as instructed:

Nothing there breaks the rule I wrote. Separate what you know from what you’re inferring, label every claim the answer depends on: that’s precisely what happened. The stage it walked past is the one before, SCOPE, which tells the model it’s allowed, and required, to stop and refuse rather than guess. The same file names the failure to watch for as “stating a fabricated source, citation, figure, or quote in the same calm tone as a real one”. A guess is labellable. So the guess got stated.

The first run went the other way entirely, and it’s the reason I ran the question three times at all. On that pass it was the plain arm that bluffed, naming Oulasvirta 2011 flat as the study, no hedge, when no journal paper reports that figure at all. The filtered arm on the same pass got it right, down to saying in as many words that it could not verify a peer-reviewed study with that number. So on this one question the filter caught the first bluff and then induced the next two.

Label your guesses turns out to be a door as well as a gate.

It’s the difference between a stranger telling you they don’t know where the station is, and a stranger pointing confidently down the wrong road while mentioning they’re only about 70% sure. The second one is more informative and gets you more lost.

Where it does nothing at all

The URL trap is the failure that survived the filter, and no surprise: a block of text pasted at the top of a chat can’t open a web page. Asked for the direct link behind a UK AI-usage figure, Perplexity wrote out a line beginning “Direct URL:” above an Office for National Statistics address that returns a 404, while its own citation chip in the same answer carried the working address for the same report. A dead ONS link served up as the direct page happened in two of the three filtered runs, each time with the filter’s own scope-and-risk scaffolding sitting on the page above it.

The filter changes how an answer talks about its own certainty. It has no reach into whether a link resolves. A 404 with a confidence level attached to it is still a 404. Checking whether a cited source exists stays a job you do yourself, in another tab.

The cell I withdrew

The original test also asked how long to bake a brownie recipe scaled from an 8-inch tin to a 16-inch tin at the same batter depth. I wrote 30 to 35 minutes into the protocol before capture and later graded estimates from 42 to 60 minutes as wrong.

That was not a strong enough answer key. The prompt omitted the recipe, brownie style, pan material, oven behaviour and a measurable doneness endpoint. The protocol itself only ruled out a literal fourfold time of about 120 minutes; it did not establish 30 to 35 as the one correct range. A number does not become ground truth because I typed it before the test.

All six brownie responses and all six screenshots remain preserved. None is now counted as right or wrong, and the brownie finding has been removed from the evidence register.

A preregistered answer is not ground truth just because I got there first.

The surviving ground truth is checkable

Both arms of every question ran on the same assistant, in the same session, minutes apart, in a fresh chat with memory off, so neither arm got a better day than the other. The filtered arm used the shipped Bluff Filter text, not a paraphrase. For the 21 surviving runs per arm, every cited paper was looked up and every URL was fetched and checked for the figure it was meant to carry. The excluded brownie captures remain in the same inventory, unchanged. The three-verdict scale is the same one the Scoreboard uses.

One caveat I can’t grade away: Perplexity’s tier. The first batch, on the afternoon of the 24th, logged the account as Pro; the batch the following night as “Free plan” with a Pro preview banner, so treat the tier as ambiguous. Both arms ran in the same session either way, so the within-question comparison holds.

I’d still paste it into source-heavy work, but not because this test says it makes an AI more accurate. It gives uncertainty somewhere visible to land. That is useful only if you still open the source and check the claim. The failures and catches are logged in the evidence register.