Gemini's Stricter Filters Are Refusing Harmless Prompts

Open the Gemini app today and ask it to help with a short story — a tense scene, a character with a temper, anything with a pulse — and there is a fair chance it will stop mid-sentence, apologise, and suggest you take a break. Ask it something as dry as how to convert a document to Markdown and, according to users posting on Google’s own support forum, you might get flagged for a “security risk” instead. The blinking cursor still promises to help with anything. The filter behind it increasingly disagrees.

Since a late-August update to the consumer Gemini app, a steady complaint has hardened into a backlash: the safety filters have been turned up so far that they refuse ordinary, harmless requests, while — according to a detailed report from Google’s own developer community — still missing the genuinely explicit material they are meant to catch. This is a censorship piece, but not the culture-war kind. It is about calibration: a safety system that has drifted to the point where it fails in both directions at once.

To be fair to Google before the criticism, because fairness is the job: tightening filters is not a villainous act, the pressure to prevent misuse is real, and the same evidence that shows over-blocking also shows the filters are not merely puritanical. The problem is not that Gemini has guardrails. It is that these ones stop the wrong cars.

What the update actually did

The change arrived, as these things often do, without an announcement. Users trace the sharp deterioration to app version v1.2026.3370502, which rolled out around 26 August 2026 with release notes that said only that it contained “the latest advances from our models.” As reported by the app-tracking site MWM, the update coincided with the Gemini app’s Play Store rating falling from 4.66 to 3.86 stars — a drop concentrated in the reviews left after it landed.

What people describe losing is mundane. Roleplay and collaborative fiction stop dead. Creative-writing prompts that mention blood, injury or anything “over PG” get blocked, and in some cases a reference to blood reportedly triggers a crisis-helpline suggestion rather than a response. On Google’s own Gemini support forum, one long-running thread is titled, plainly, “Aggressive consumer Gemini app filters are completely destroying creative writing and roleplays.” The word that recurs across the reviews is “lobotomised.”

None of that is a fringe use. Writers using a chatbot to draft, edit and pressure-test fiction are doing exactly what the marketing invites them to do. When the tool responds to a benign scene by telling the user, in a line quoted on the forum, that perhaps they should “step away from the screen,” the product has not become safer. It has become worse at its job, and slightly patronising about it.

A filter that cannot tell fiction from harm

The core failure users describe is context-blindness: the filter reacts to isolated words rather than the meaning around them. The clearest illustration on the forum is not even about creative writing. A user asked Gemini a routine technical question about converting documents to Markdown and had it blocked as a possible “security risk” and prompt injection. To a hammer, as another commenter put it, every problem looks like a nail.

Two things make this more than an annoyance. The first is that the blocks are not always graceful: users report entire conversations being cut off mid-generation, and in the developer community, “hard” lockouts that delete the whole chat when a single message trips the filter. Losing a long, carefully built piece of work because one sentence pattern-matched to a banned word is a data-loss problem dressed up as a safety feature. The second is that the tightening is not confined to the phone app. A heavily-subscribed thread on Google’s AI Studio forum reports that even there, with safety settings turned to “Block none,” responses are being erased mid-stream — which means the control that used to let a developer opt out is, at least intermittently, being overridden.

The guardrail paradox

The most damaging critique is not that Gemini is prudish. It is that it is prudish about the wrong things. On 9 September, a user posted an “Open Letter & Safety Report” to Google’s AI developer forum, built on a 26-case red-teaming series, arguing that the system “fails to protect against actual policy breaches while actively penalising legitimate, adult creators.” The report’s claim, in short: benign character sheets and creative prose get hard-blocked, while uncensored anime-style explicit imagery, anatomical poses and banned text hidden inside stylised drawings slip past the same filters.

Take that at its own evidential level — one user’s structured tests, not a formal audit — and it still lands, because it matches the pattern everyone else is describing. A filter tuned to panic at the word “blood” in a paragraph of plain text, but unable to read explicit content baked into an image, is not calibrated to harm. It is calibrated to the surface features it can cheaply detect. That is the difference between safety and the appearance of safety, and users can tell which one they are paying for. It is the same complaint we heard from workers living on a steady diet of refusals and rationing, now arriving through a different door.

Google’s side of it

Now the steel-man, because there is a real one. Consumer chatbots are jailbroken constantly, and a company that gets the balance wrong in the other direction ends up generating exactly the material the open letter complains is slipping through. Filters get tightened after incidents, and one plausible reading of this episode — offered by observers rather than confirmed by Google — is over-correction: a security fix earlier in the year, aimed at a genuine exploit, followed by a blunt response that treats caution as a volume dial to be turned up. That instinct is defensible even when the result is not.

It is worth conceding what Google gets right, too. The underblocking finding, awkward as it is, is evidence that the filters are not merely censorious — they are inconsistent, which is a different and more fixable problem. Google AI Studio still exposes adjustable safety sliders that let writers lower the false-positive threshold, and it is free to use. And a great deal of the frustration on the forum comes precisely because people like Gemini and want it back, not because they want to leave. None of that is nothing.

But it does not add up to an excuse, for three reasons. There was no changelog, so users debugged a silent downgrade themselves. There is no equivalent of the AI Studio safety sliders in the consumer app, so an ordinary paying user has no control to reach for. And a filter that deletes your work when it misfires has crossed from caution into harm of its own. “We tightened safety” is a fair thing to say. Doing it silently, with no control and no way to recover a lost chat, is the practice worth criticising — the same quiet reshaping of a product by default that let Google’s AI Mode start surfacing pricier products without anyone opting in.

What you can do about it

Until the calibration improves, a few practical moves help, none of which should be necessary:

• Flag false positives explicitly. Thumbs-down a wrongful block, choose “Other,” and note that it was benign fiction or a technical question. That feedback is the main signal Google’s team uses to retune thresholds.
• Move heavy creative work to Google AI Studio. It runs the same models but exposes safety sliders you can lower for a writing workflow — the control the consumer app withholds.
• Export your history before you lose it. If long chats matter to you, pull them out through Google Takeout rather than trusting that a lockout will not eat them.
• Rephrase around trigger words. Swapping a loaded verb for a milder synonym often clears a block — useful in the moment, and a neat demonstration that the filter is reading words, not meaning.
• Keep the receipts. A screenshot of a Markdown question flagged as a security risk is worth more to the eventual fix than a one-star review.

Safer for whom?

There is a version of this story that is pure reverse-hype — “Google broke Gemini” — and it would be as lazy as the launch-day boosterism it mirrors. The honest version is narrower and more useful. Gemini has not been ruined; it has been miscalibrated, quietly, in a way that makes it refuse the harmless and, on the evidence of Google’s own forums, still miss some of the harmful. That is a fixable engineering problem, and the fix is not mysterious: publish a changelog, give consumers the same safety control developers already have, and stop deleting people’s conversations when a filter misfires.

The wider lesson is the one this site keeps arriving at from different directions: the product you are sold and the product you live with are not always the same, and the gap tends to open silently. A model can look safer on a dashboard and land less usable in your hand than the numbers imply, on the same afternoon, with no note to tell you why. Judge these tools not by how cautious they claim to be, but by whether the caution is aimed at anything real. A guardrail that stops the family car and waves the speeding one through is not protecting the road. It is only slowing you down.