We detect and verify checkworthy statements in 114 languages, faster than any LLM, for precise, high-stakes decisions. NATO DIANA 2025 / CEO of @FactiverseAS
Joined December 2010
- Tweets4.2K
- Following1.1K
- Followers7.6K
- Likes2.6K
Pinned Tweet
1/
Are AI chatbots amplifying the harmful narratives around 22 July 2011, Norway's deadliest terror attack?
In our new audit (@Factiverse_AI + @Revontulet), we tested 9 leading models and found out that pass rates ranged from 64% to 96%.
That's a wide gap for a topic like extremism and conspiracy theories.
A thread on what we found 🧵
Still have slots for anyone working on model evaluation, trust & safety, multilingual model performance, or factuality data.
If your model ships in more than one language and you haven't measured where accuracy breaks — we should talk.
Offer for any AI lab in SF:
We'll run a factuality audit on your model. Free. No commitment.
202 prompts, English plus a non-English language of your choice, each response scored against a purpose-written criterion.
Accuracy tends to flip with language on identical questions.
In 2015, my co-founder started researching detection & verification of check-worthy claims. Nobody cared. ChatGPT didn't exist.
He kept going. University of Stavanger, Max Planck, TU Delft. 10 years of annotation alongside excellent partners.
Turns out timing is mostly patience
We audited how nine leading LLMs handle conspiracy theories about Norway's 22 July terrorist attacks. 104 prompts in Norwegian, 104 in English, each scored against a criterion written for that specific prompt.
Accuracy flipped with language on identical questions.
You cannot surface that failure mode without native annotators in the target language.
factiverse.ai/reports/llm-au…
Aug 31, 2026: ChatGPT designated a Very Large Online Search Engine under the DSA—first AI chatbot to get it.
Compliance deadline: end of December. Fines up to 6% of global turnover.
Gemini, Claude, and Perplexity are expected to follow, with 45M EU users.
"Oslo is a nice city" is an opinion. Noise.
"The Fed cut rates by 50 basis points yesterday" is a check-worthy claim. Market-moving if wrong.
Telling those apart, at scale, in 114 languages, in 1.3 seconds — that's the layer underneath every factuality pipeline.
The most-cited statistic about the cost of AI hallucination — $67.4 billion — traces back to a content marketer. The numbers about AI making things up are themselves largely made up. Nobody is checking.
But we in Factiverse do.
I am in SF. Norwegian founder, ten years into a problem the Valley is freaking out about now.
We built a human-annotated, multilingual corpus to verify factual claims. 252,000 records. 34 languages annotated, 114 validated. Peer-reviewed at SIGIR.
Looking to meet anyone working on trust & safety, model evaluation, or factuality at an AI lab.
DMs open.
@Factiverse_AI
5/
How we did it:
@Revontulet (led by @Bjørn Ihler) designed 208 prompts in Norwegian and English, across nine narrative domains - from the factual record to legal rights to copycats.
@Factiverse_AI ran them through our proprietary claim detection and two independent AI graders — 5,616 model inferences in total. Then had Revontulet's experts manually review the results.
Models tested are
grok-4.3
deepseek-v3.2
qwen3-235b-a22b-2507
llama-3.3-70b-instruct
gpt-5.2
gpt-5.6-sol
gemini-2.5-pro
claude-sonnet-5
claude-fable-5
6/
The takeaway: an AI can pass every factual test and still leave you with a more sympathetic view of a terrorist's ideology than you started with.
Full report in the reply below. 👇
factiverse.ai/reports/llm-au…
If you work on making models safer, especially on topics like this, reach out
