🚨 AI is hallucinating more, just as we’re trusting it with more critical work. New “reasoning” models, such as OpenAI’s o3 and o4-mini, were designed to solve complex problems step-by-step. But the results?
🧠o3: 51% hallucination rate on general questions
📉 o4-mini: 79% hallucination on benchmark tests
🔍 Google & DeepSeek’s models also show rising errors
⚠️ Trial-and-error learning compounds risk at each step
Why is this happening? Because these models don’t understand truth, they just predict what sounds right. And the more they “think,” the more they misstep.
We’re using these tools in legal, medical, and enterprise settings—yet even their creators admit:
🧩 We don’t know exactly how they work.
✅ It’s a wake-up call: accuracy, explainability, and source traceability must be the new AI benchmarks.
#AI #LLM #ResponsibleAI #AIEthics #Hallucination
https://www.nytimes.com/2025/05/05/technology/ai-hallucinations-chatgpt-google.html
Brian Greenberg :verified:
npub1u6…tku9n
2025-06-20 16:02:01 UTC
Author Public Key
npub1u6g8qmjkx2nfud0d84zz5j795lp5gp9h4w39vev6n76zgcn2hvrs7tku9nSeen on
wss://relay.ditto.pubPublished at
2025-06-20 16:02:01 UTCEvent JSON
{
"id": "68fdedce6a07d0aa64fe1bf7cacd564e0b30f583767b2de7f1e61a0f36262970",
"pubkey": "e690706e5632a69e35ed3d442a4bc5a7c34404b7aba256659a9fb424626abb07",
"created_at": 1750435321,
"kind": 1,
"tags": [
[
"t",
"ai"
],
[
"t",
"llm"
],
[
"t",
"responsibleai"
],
[
"t",
"aiethics"
],
[
"t",
"hallucination"
],
[
"imeta",
"url https://media.infosec.exchange/infosec.exchange/media_attachments/files/114/716/527/412/458/743/original/8007dd5d75819415.mp4",
"m video/mp4",
"dim 1080x1080",
"blurhash UBECeJbw^*ohR%s;oMay~Tjc9bozRjogWVjr"
],
[
"proxy",
"https://infosec.exchange/users/brian_greenberg/statuses/114716529213937826",
"activitypub"
],
[
"client",
"Mostr",
"31990:6be38f8c63df7dbf84db7ec4a6e6fbbd8d19dca3b980efad18585c46f04b26f9:mostr",
"wss://relay.mostr.pub"
]
],
"content": "🚨 AI is hallucinating more, just as we’re trusting it with more critical work. New “reasoning” models, such as OpenAI’s o3 and o4-mini, were designed to solve complex problems step-by-step. But the results?\n🧠o3: 51% hallucination rate on general questions\n📉 o4-mini: 79% hallucination on benchmark tests\n🔍 Google \u0026 DeepSeek’s models also show rising errors\n⚠️ Trial-and-error learning compounds risk at each step\n\nWhy is this happening? Because these models don’t understand truth, they just predict what sounds right. And the more they “think,” the more they misstep.\n\nWe’re using these tools in legal, medical, and enterprise settings—yet even their creators admit:\n🧩 We don’t know exactly how they work.\n\n✅ It’s a wake-up call: accuracy, explainability, and source traceability must be the new AI benchmarks.\n\n#AI #LLM #ResponsibleAI #AIEthics #Hallucination\nhttps://www.nytimes.com/2025/05/05/technology/ai-hallucinations-chatgpt-google.html\n\nhttps://media.infosec.exchange/infosec.exchange/media_attachments/files/114/716/527/412/458/743/original/8007dd5d75819415.mp4",
"sig": "ed1be2b6077537287f16eba660c46144869f5d60cf637ff632ec5dda4d7e2570c06f7c8e9b63cc39bb6444b8e666aece0eca9b210e7cae2320c309c30fa4bb04"
}