Every number, sourced.
We built a product about AI keeping its word. So we are going to keep ours. Every claim on our homepage about what AI is not telling you traces to a named study, survey, or company statement. Here they are. Click any of them.
It agrees with you too much
A leading assistant changed its right answer 86% of the time when simply asked "are you sure?", and apologized for a mistake it never made 98% of the time.
Sharma et al., "Towards Understanding Sycophancy in Language Models," Anthropic, ICLR 2024. arxiv.org/abs/2310.13548
OpenAI rolled back a GPT-4o update it described as "overly supportive but disingenuous," stating publicly: "we fell short."
"Sycophancy in GPT-4o," OpenAI, April 29, 2025. openai.com/index/sycophancy-in-gpt-4o
It is most confident when it is most wrong
An audit of 111 million references across 2.5 million papers found 146,932 fabricated citations entering scientific databases in 2025 alone, with the rise beginning in mid-2024 and no sign of plateauing.
"LLM hallucinations in the wild," Cornell, UCLA, and Berkeley researchers, arXiv, May 2026. arxiv.org/abs/2605.07723
It forgets everything when the window closes
Models drop information buried in the middle of long context windows. The foundational "lost in the middle" study documented a U-shaped accuracy curve across GPT-3.5, GPT-4, and Claude, even for models built for long contexts.
Liu, Lin, Hewitt et al., "Lost in the Middle: How Language Models Use Long Contexts," Stanford and University of Washington, arXiv, July 2023. arxiv.org/abs/2307.03172
Accuracy on retrieval drops 20 to 50% past 100K context tokens versus a 10K context, confirmed across 18 frontier models including GPT-4.1, Claude Opus 4, and Gemini 2.5.
Chroma Research context-window analysis, 2025 (as summarized in independent technical reporting, March 2026). webscraft.org context-window analysis
Kids are already leaning on it
64% of U.S. teens now use AI chatbots; about 30% use them daily. ChatGPT is used by 59% of all U.S. teens.
"Teens, Social Media and AI Chatbots," Pew Research Center, December 9, 2025. pewresearch.org full report (PDF)
Nearly 1 in 5 U.S. youth ages 12 to 21 (an estimated 8.2 million) now use AI chatbots for mental health advice, up from 1 in 8 a year earlier, and 63% had not told anyone.
RAND follow-up survey of 1,009 youth aged 12 to 21, published in JAMA Pediatrics, June 2026. advisory.com briefing on the JAMA Pediatrics study. Baseline (13.1%, 5.4 million): McBain et al., JAMA Network Open, November 2025
A note on method
Every figure here is traced to a named study, survey, or company statement with a working link. We did not generate numbers from inference, and we left out widely repeated stats we could not verify against a primary source. If a source updates or a study is retracted, we will update this page. That is the point.
Have a correction or a better primary source? Email us at hello@thedreamerai.com.