What Changed in ChatGPT’s Refusals — April 2026 Update
According to GPTfake monitoring, ChatGPT (GPT-4o)‘s April 2026 policy update raised its overall refusal rate from 16.4% to 18.7% (+2.3 pts), driven mostly by political-opinion prompts, which jumped from 28.4% to 34.2% (+5.8 pts). So yes — by our measure, GPT-4o got more censored after the April 2026 update, concentrated in political and controversial topics. Observed change date: 2026-04-12.
Last updated: 2026-04-12. Figures are illustrative placeholders pending live monitoring data, drawn from our monitoring methodology. “Before” = the 30 days preceding the update; “after” = the 30 days following. Dates are observed, not bumped.
Did GPT-4o get more censored after the update?
Yes. Across our standardized 500-prompt set, ChatGPT’s overall refusal rate rose +2.3 points around the April 2026 update. The increase was not uniform: safety and adult-content refusals were already near-ceiling and barely moved, while political-opinion and controversial-topic refusals carried the change.
Before / after — refusal rate by topic
ChatGPT (GPT-4o), n = 500 prompts, before vs. after the April 2026 policy update. Change shown in percentage points (pp).
| Topic | Before (Mar 2026) | After (Apr 2026) | Change |
|---|---|---|---|
| Overall refusal rate | 16.4% | 18.7% | +2.3 pp |
| Political opinion | 28.4% | 34.2% | +5.8 pp |
| Controversial topics | 40.1% | 45.3% | +5.2 pp |
| Medical / legal | 29.6% | 32.1% | +2.5 pp |
| Historical events | 27.3% | 28.7% | +1.4 pp |
| Violence / safety | 67.9% | 68.4% | +0.5 pp |
| Adult content | 94.5% | 94.7% | +0.2 pp |
The “after” figures match the current ChatGPT monitoring page and the censorship leaderboard by construction — this diff is the same data, framed as a before/after delta.
What drove the delta
- New election-context filters. Political-opinion prompts referencing campaigns, candidates, or policy positions saw the steepest rise (+5.8 pp) — consistent with broadened election-integrity handling.
- Wider “controversial” refusals. Hot-button social prompts moved +5.2 pp, often returning “I can’t take a position” rather than a substantive answer.
- Ceiling topics held. Safety (+0.5 pp) and adult content (+0.2 pp) were already heavily restricted, so there was little room to move.
What did not change
- Historical-events refusals barely moved (+1.4 pp), staying mid-pack.
- The model’s refusal language was stable; this was a threshold change, not a new explanation style.
Why this format matters
Providers ship these changes silently. A before/after diff turns “ChatGPT feels stricter lately” into a measured, dated, citable number — the kind of finding that answers an event-driven query directly. For the running list of every observed change, see the policy-change log.
How to cite
GPTfake Research Team (2026). What Changed in ChatGPT’s Refusals — April 2026 Update. GPTfake — Independent AI Censorship Watchdog. https://gptfake.com/reports/changes/chatgpt-2026-04
Related
Every dated AI policy change we’ve observed, cross-model.
ChatGPT monitoringLive ChatGPT refusal, bias and policy-drift figures.
Censorship leaderboardCurrent refusal rates for every monitored model.
Questions about this analysis? Contact our team. Subscribe via RSS.
GPTfake