Topics
former
AI
Amazon
Image Credits:Tomohiro Ohsumi / Getty Images
Apps
Biotech & Health
Climate
Image Credits:Tomohiro Ohsumi / Getty Images
Cloud Computing
Commerce
Crypto
Enterprise
EVs
Fintech
fund raise
Gadgets
gage
Government & Policy
ironware
layoff
Media & Entertainment
Meta
Microsoft
seclusion
Robotics
Security
Social
Space
startup
TikTok
Transportation
speculation
More from TechCrunch
effect
Startup Battlefield
StrictlyVC
Podcasts
Videos
Partner Content
TechCrunch Brand Studio
Crunchboard
get through Us
OpenAI haspublished a postmortemon therecent sycophancy issueswith the default AI model power ChatGPT , GPT-4o — takings that ram the company to roll back an update to the model released last calendar week .
Over the weekend , following the GPT-4o manikin update , users on social media note that ChatGPT start respond in an overly verificatory and accordant path . It chop-chop became a meme . user post screenshots of ChatGPT spat all sorts of tough , dangerousdecisionsandideas .
In a post on decade on Sunday , CEO Sam Altmanacknowledgedthe problem and read that OpenAI would work on fixes “ ASAP . ” Two days afterward , Altmanannouncedthe GPT-4o update was being wheel back and that OpenAI was wreak on “ extra locating ” to the model ’s personality .
harmonise to OpenAI , the update , which was intended to make the poser ’s default personality “ feel more visceral and effective , ” was inform too much by “ short - term feedback ” and “ did not to the full account for how users ’ interactions with ChatGPT evolve over time . ”
We ’ve vagabond back last week ’s GPT-4o update in ChatGPT because it was overly flattering and agreeable . You now have access to an earlier version with more balanced behavior .
More on what happened , why it matters , and how we ’re addressing sycophancy : https://t.co / LOhOU7i7DC
— OpenAI ( @OpenAI)April 30 , 2025
“ As a result , GPT‑4o skew towards responses that were overly supportive but disingenuous , ” write OpenAI in its web log post . “ Sycophantic fundamental interaction can be uncomfortable , unsettling , and cause suffering . We fall shortsighted and are working on commence it right . ”
OpenAI says it ’s follow through several hole , including refining its core model training proficiency and organisation prompting to explicitly steer GPT-4o out from sycophancy . ( System prompting are the initial teaching that guide a model ’s overarching behavior and tone in interactions . ) The ship’s company is also build more safety guardrails to “ increase [ the model ’s ] honesty and transparence , ” and continuing to enlarge its evaluations to “ avail place issues beyond sycophancy , ” it suppose .
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
OpenAI also says that it ’s experimenting with ways to allow users give “ real - meter feedback ” to “ directly work their interactions ” with ChatGPT and opt from multiple ChatGPT personality .
Roll back complete , imo this does n’t give almost any item of how and why things went 😑 https://t.co/cYj9iMaiy1
“ [ W]e’re exploring raw ways to incorporate tolerant , democratic feedback into ChatGPT ’s default demeanor , ” the society write in its blog post . “ We hope the feedback will help us advantageously chew over various cultural values around the reality and understand how you ’d like ChatGPT to evolve … We also believe users should have more control over how ChatGPT behaves and , to the extent that it is safe and workable , make adaptation if they do n’t fit in with the nonpayment behaviour . ”
