Say No to Easter Eggs

Say No to Easter Eggs: Planted Errors in AI Police Reports

Some AI tools plant fake words in police report drafts to prove officers read them. Why that tests attention, not accuracy, and what to demand instead.

Ed Choudhry

Research & Policy Series

Some AI report tools plant odd words in a draft to prove an officer read it. It is the wrong safeguard for the wrong failure.

At P3 in Santa Monica, I said it on stage: say no to Easter eggs. Great for games. Not for reporting. This is the long version.

An Easter egg in an AI-drafted police report is a deliberate insertion: an odd word or sentence the officer must find and delete before the report moves forward. Vendors offer it as proof of human review. The intent is reasonable. The design is not.

An Easter egg tests whether an officer noticed a planted word. It does nothing about the errors that actually threaten a case: the plausible ones a system writes when it does not have enough context. A police report is sworn testimony, not a scavenger hunt. Nothing false belongs in a draft, including things put there on purpose.

Where Easter eggs come from

Easter eggs began as a gift from developers to players. They were never meant for serious systems.

The earliest known video game Easter egg is in Moonlander (1973). Fly far enough sideways and you find a McDonald's on the Moon. Land next to it, and your astronaut walks over for a meal.

The name came later. Warren Robinett built Adventure for the Atari 2600, developed in 1979 and released in 1980. Atari did not credit its programmers, so Robinett hid a secret room with the words "Created by Warren Robinett." A teenage player found it and wrote to Atari. Management considered pulling it. Steve Wright, Atari's director of software development, argued to keep it and called hidden features like it Easter eggs, for players to find.

The tradition grew from there. The original Mortal Kombat (1992) hid Reptile, the first secret character in fighting game history. Mortal Kombat II added sound designer Dan Forden popping into the corner to shout "Toasty!" Call of Duty: World at War (2008) hid Nazi Zombies behind the campaign. That bonus became one of the franchise's most popular modes, its maps full of Easter egg quests.

Every one of those works for the same reason: the stakes are play. The player chose to hunt, and finding the egg is the reward.

Serious software went the other way. By 2005, Microsoft's operating system division had a "no Easter eggs" policy under its Trustworthy Computing initiative, and one of its engineers called them irresponsible. Security writers warned that hidden, untested code has no place where confidential data lives. If the software industry decided hidden content does not belong in trustworthy systems, it does not belong in a document that can send someone to jail.

Apple drew the same line. Steve Jobs banned Easter eggs from Apple products when he returned to the company.

Planted fakes have a history too

Easter eggs have a more serious cousin: the deliberate fake, planted to catch someone. Every version that worked kept the fake out of anything that mattered.

Mapmakers plant "trap streets" to catch copycats. In the 1930s, the General Drafting Company invented Agloe, New York, an anagram of its founders' initials, at an empty Catskills crossroads. Years later, a general store opened on that spot and took the name from the map. A rival's map then listed Agloe as real. The planted falsehood had escaped into the record.

Reference works do it too. The 1975 New Columbia Encyclopedia carried an entry for Lillian Virginia Mountweazel, a person who never existed. Fake entries like hers are still called mountweazels.

Software engineers have "seeded" known bugs since the 1970s, starting with Harlan Mills at IBM, to measure how many real bugs their testing misses. Steve McConnell, author of Code Complete, warned that a common failure is forgetting to remove seeded defects, and said to remove every one before release.

The lesson is consistent. Fakes are tools for testing copiers, testers and systems, and they stay out of the finished product. A police report is the finished product.

How Easter eggs got into police reports

In police reporting, an Easter egg is a review check: plant something wrong, and make the officer find it before signing. The practice is spreading.

The largest vendor offers it. KSTP's 5 Investigates reported in 2025 that Axon's Draft One, which drafts reports from body-camera audio, includes an option to add Easter eggs so drafts get reviewed.

A startup builds it in. In July 2026, StateScoop described Code Four scattering obscure words such as "axolotl" through AI-drafted affidavits and case files. Officers must delete them before finalizing, and the company does not tell them which words to look for. Its CEO says this forces meaningful edits instead of copy and paste.

A city now relies on it. This October, Phoenix began a 90-day pilot of an audio-only drafting tool, running through December 31, 2026. A city councilmember who championed the pilot told ABC15 that "there actually is an error placed in the report that the officer must find" before signing the attestation. The station reported the error is embedded in every draft.

Read that again. Every draft of a sworn report now begins with a known falsehood in it, put there on purpose.

The intent behind all three is good. Everyone in this field agrees the risk is rubber-stamping: an exhausted officer signing a fluent draft without reading it. The question is whether planting errors is the right answer. We think it is not.

Where hallucinations come from

A hallucination is a plausible but false statement, and it starts where the information runs out.

OpenAI's own researchers put it plainly in 2025: models hallucinate because standard training and evaluation reward guessing over admitting uncertainty. They also note some questions cannot be answered at all, because the information is unavailable. A model graded on sounding right will guess.

Now apply that to a police report. An audio-only system works from a transcript. It hears words; it does not see the scene. Which hand moved, what was handed over, what color the light was, how many people stood there: if nobody said it out loud, it is not in the input. The draft still has to read as a complete narrative, so the gaps get filled with the most plausible guess.

We laid out the evidence in our multimodal vs. audio-only article. In Heber City, Utah, a body camera picked up a Disney movie, and an AI draft said the officer turned into a frog. Anchorage police dropped an audio-only tool after officers found it missed what they saw but never said aloud. Our guide to AI-assisted report writing also walks through a leading vendor's own published paper: an audio-only pipeline that processes no video and leaves placeholders wherever facts cannot be determined.

What temperature does, and doesn't, fix

Vendors often say they keep "temperature" low to avoid embellishment. Temperature is a sampling setting: higher values make output more varied, lower values make the model pick its most likely next words. Research does show hallucination rates rise with temperature. But setting it near zero does not stop them. A 2026 benchmark found temperature's effect on hallucination negligible; the differences came from how well models were grounded.

That is the point. A low-temperature model with a gap in its input does not leave the gap empty. It fills it with the most likely-sounding answer, stated confidently. Turning the dial down makes the guess more predictable, not more true. The fix is more context, not less creativity.

An Easter egg sits on top of this problem and never touches it. The planted word is in the draft. The guessed fact is in the draft too. Only one of them looks wrong.

From recall to scavenger hunt

Reviewing a police report is a recall task. An Easter egg turns it into a search task, and decades of research on search say that is the wrong trade.

Two different jobs

A real review means replaying the incident and checking every line against memory and footage: slow, deliberate, about meaning. Hunting for a planted word is visual search: scanning the surface for something that does not belong. Readers measurably change how they process words depending on which job they think they are doing. Eye-tracking work by Schotter and colleagues (2014) found people reading for comprehension and people proofreading for errors process the same text differently. Tell an officer the task is "find the egg," and that is the task their eyes perform.

Obvious targets are the easy ones

Not all errors are equally visible. In a 2022 eye-tracking study, Burgoyne and colleagues found readers caught misspellings more often than wrong words, real words used in the wrong place. A word like "axolotl" belongs to the easy category: out of place on its face. A hallucinated fact belongs to the hard one: real words, plausible, wrong. The Easter egg trains officers on the error that was never the threat.

Finding one makes you miss the next

Radiology has studied this for more than 50 years under the name satisfaction of search: after finding one abnormality, a reader is more likely to miss a second. The effect shows up in proofreading too. In an eye-tracking study at the University at Albany, readers who found an easy typo first searched less thoroughly afterward and were less likely to catch a subtle one. An Easter egg guarantees an easy first find in every report. The officer finds it, deletes it, and the review feels complete.

Aviation already learned this lesson

The strongest case for planted errors comes from airport security. Screeners see fictional threats projected onto a small share of X-ray images, about 1% to 4% of bags, so their performance can be measured and rare threats stay on their radar. Two details matter. First, the fake must be realistic: researchers report screeners sometimes spot projected threats because they look artificial, which defeats the purpose. Second, the bag never changes; the fake exists only on the screen and in the log. Easter eggs break both rules. They are deliberately unrealistic, and they live inside the document itself.

Reading a draft can rewrite the memory

There is a deeper problem than missed errors. Psychologist Elizabeth Loftus spent 30 years documenting the misinformation effect: exposure to misleading information after an event impairs memory of the event itself. An officer who reads a fluent AI draft before recalling the scene is receiving post-event information, accurate or not. Our guide cites a 2024 study by Agudo and colleagues in which people caught 36.8% of AI errors when they saw the AI's answer first, and 66.2% when they gave their own judgment first. Same people, same errors; only the order changed.

That is the difference between memory enhancement and memory contamination. An Easter egg does nothing about order. It only adds one more thing to read.

An Easter egg proves an officer found a word. It cannot prove the officer remembered the scene.

Why Easter eggs fail

Easter eggs fail in four ways, and none of them is about intent.

They treat the symptom. The disease is a system that guesses when it cannot see. Rubber-stamping is the symptom. A planted word tests the officer and leaves the gap exactly where it was.

They miss the errors that matter. Real errors are plausible. Our guide describes a King County, Washington, prosecutors' memo flagging an AI-assisted report that placed an officer at a scene he never attended. It passed officer review. The office then told prosecutors not to accept AI-drafted narratives. No Easter egg would have caught that error, because it did not look wrong.

They are interchangeable with hallucinations. On the page, a planted line and a fabricated line look the same. Once a vendor admits its system inserts false content on purpose, every odd sentence in every report invites the question: was this put here, or did the model make it up? That is a gift to anyone cross-examining the officer.

They get preserved. California's SB 524 took effect January 1, 2026. As our state-by-state tracker explains, it requires agencies to keep the first AI-generated draft for as long as the final report, with an audit trail. In California, the planted error is now part of the retained record.

There is one more problem. Where Easter eggs are an option, they can be turned off. Our guide notes reporting that some agencies have quietly disabled the safeguards built into AI reporting tools. A safeguard that can be switched off is a setting, not a guardrail.

What we build instead

The alternative to a trap is a real question. We built C.O.D.E. 10-8 around the cause, not the symptom.

Context first. Our system is multimodal. It works from far more of the evidence than a transcript, so fewer facts depend on what someone happened to say out loud. Fewer gaps means fewer guesses.

Gaps surfaced, not filled. Where the evidence cannot establish something, our system is designed to say so, not to smooth it over with a plausible sentence.

The officer's memory leads. The workflow is built so the officer engages with the incident as its author, supplying what only they know, instead of auditing a finished narrative for planted mistakes.

No planted words. No traps. Nothing false in the draft, ever.

We would rather show you than describe it. If your agency is weighing AI for report writing, book a demo and we will walk you through exactly what C.O.D.E. 10-8 does when a draft reaches something it cannot know.

The difference fits in two lines.

A trap asks: are you paying attention? A flag asks: what do you know that the camera doesn't?

We are honest about the limits. Even multimodal AI cannot smell alcohol, read a suspect's history with an officer, or articulate probable cause. As our guide puts it, AI drafts what the camera captured. The officer provides everything else.

What to ask before you sign

If your agency is evaluating AI for report writing, ask every vendor these six questions, including us:

  1. Does it see the evidence, or only hear it?

  2. Can every sentence be traced to a source?

  3. Does it ever plant content on purpose?

  4. When it doesn't know, does it flag or fill?

  5. Can anything reach the record without the officer?

  6. Is "no training on our data" in the contract, not just the pitch?

Easter eggs belong in Adventure, Mortal Kombat and Call of Duty, where the player chose the hunt and the prize is the fun. A police report is not a game. It is sworn testimony that can take away someone's liberty. Nothing false belongs in it, including things put there on purpose.

The agencies that lead in this space won't be the ones that adopted AI first. They'll be the ones that adopted it most responsibly. Say no to Easter eggs.

Sources

Barricade AI research

Easter eggs in police reporting

History of Easter eggs

Planted fakes

Hallucinations and temperature

Attention, search and memory

Barricade.ai // All rights reserved. 2026

Privacy Policy

Terms of Use