Breaking
Hong Kong’s 5-year plan consultation has sparked valuable discussionsClop builds custom web shell for Windchill servers in data theft campaignTop Chef turns to tears as host Kristen Kish redefines the roleMicrosoft Copilot Personal Flaws Could Let One Click Exfiltrate Data From Connected AppsMLflow flaw exploited to steal cloud credentialsESMA consults on reporting framework for clearing activity at recognised third-country CCPsECB collateral surprises cut bank risk and narrow bond spreadsFlores power restored after earthquake as all 1.144m customers back onlineAI firms undercount personal use of their tools as scepticism grows over dataCursor launches rival code-hosting platform Origin as GitHub struggles with reliabilityOil prices rise to three-week high as Strait of Hormuz threat lingersJustice Department probes Andreessen Horowitz over board seats at Databricks and FivetranPeople are still unsure how AI is being used in practiceEuropean stocks fall to two-week low as bond yields rise and Middle East tensions simmerRFC 9234 adoption lags in Malaysia as route leaks threaten banks and telcosPublic Service DG Wan Ahmad Dahlan named MRT Corp chairmanPuncak Niaga sells two Selangor plots for RM237.15m to cut debtAWS to open permanent Builder Lofts in Berlin, Hyderabad and São PauloRTS Bukit Chagar station to link directly to KTM, buses and taxisThreat Brief: Mitigating Large-Scale Credential Attacks (Updated Aug. 18)Hong Kong’s 5-year plan consultation has sparked valuable discussionsClop builds custom web shell for Windchill servers in data theft campaignTop Chef turns to tears as host Kristen Kish redefines the roleMicrosoft Copilot Personal Flaws Could Let One Click Exfiltrate Data From Connected AppsMLflow flaw exploited to steal cloud credentialsESMA consults on reporting framework for clearing activity at recognised third-country CCPsECB collateral surprises cut bank risk and narrow bond spreadsFlores power restored after earthquake as all 1.144m customers back onlineAI firms undercount personal use of their tools as scepticism grows over dataCursor launches rival code-hosting platform Origin as GitHub struggles with reliabilityOil prices rise to three-week high as Strait of Hormuz threat lingersJustice Department probes Andreessen Horowitz over board seats at Databricks and FivetranPeople are still unsure how AI is being used in practiceEuropean stocks fall to two-week low as bond yields rise and Middle East tensions simmerRFC 9234 adoption lags in Malaysia as route leaks threaten banks and telcosPublic Service DG Wan Ahmad Dahlan named MRT Corp chairmanPuncak Niaga sells two Selangor plots for RM237.15m to cut debtAWS to open permanent Builder Lofts in Berlin, Hyderabad and São PauloRTS Bukit Chagar station to link directly to KTM, buses and taxisThreat Brief: Mitigating Large-Scale Credential Attacks (Updated Aug. 18)
Economy

People are still unsure how AI is being used in practice

AI companies’ own data on generative AI use is too narrow to guide policy, a new independent study shows, underscoring gaps that could affect how Malaysian regulators view the technology.

Source: MIT Technology Review · August 18, 2026 at 10:31 PM · AI-assisted report

Single-source

KUALA LUMPUR, 19 AUGUST 2026 —

Listen to this article

DomainFork Audio · read aloud

AI companies’ own data on generative AI use is too narrow to guide policy, a new independent study shows, underscoring gaps that could affect how Malaysian regulators view the technology.

Market Impact

The AI Observatory, a public research platform led by Stanford’s Trustworthy AI Research Lab, analyzed 85,633 conversational exchanges across 52 models including ChatGPT, Claude and Grok collected between 2023 and 2025. It found that nearly half of real-world conversations—48%—would be excluded by industry reports that focus on work or productivity, with those filtered-out exchanges far more likely to involve health, relationships, harassment, hate speech or sexual content.

“There is no independent source to corroborate it,” said Anka Reuel, a Stanford PhD candidate and co-lead of the AI Observatory project. “Highly consequential decisions about AI’s benefits and risks are currently being made on the basis of very limited data.”

Industry datasets such as Anthropic’s Economic Index and OpenAI’s 2025 ChatGPT report deliberately exclude non-work conversations, capturing only 30% of consumer use for work-related activity versus 52% in the AI Observatory’s broader sample. When the researchers applied Anthropic’s own filtering rules to their dataset, 48% of exchanges were removed.

The excluded data showed markedly higher rates of sensitive topics: 44.2% health and relationships versus 31.2% in Anthropic’s published analysis, adult or illicit content at 7.9% versus 2.1%, harassment and hate at 27.5% versus 5.66%, and sexual content at 16.7% versus 2.4%.

The AI Observatory also found that interaction patterns differ sharply across models and versions. Conversations on WildChat, one of the largest datasets in the study, became longer and more elaborate over time, with increases in prompt tokens, response tokens and conversation turns. Small talk rose, suggesting growing use of AI for companionship, while the models’ self-disclosure that they are chatbots declined.

Sensitive exchanges labeled by the researchers—including sexual harassment and hate speech—became less frequent over the period, which the team said may indicate improving safeguards. Topic preferences varied: Grok and Gemini were used more for information retrieval, with Grok particularly popular for news and politics but also showing higher concentrations of misinformation. Anthropic’s models were favored for coding, Google’s Gemini for social roleplay, and ChatGPT for homework assistance.

Within ChatGPT, users had shorter exchanges with GPT-3.5 and longer, more iterative conversations with GPT-4o, consistent with reports that the latter version is more engaging.

“No single company report tells the whole story,” said Shayne Longpre, a recent MIT Media Lab PhD graduate and AI Observatory co-lead.

The study aggregated conversations from seven existing datasets with user consent, totaling 24,521 exchanges across 5,000 users and 52 models. By contrast, Anthropic’s latest economic index draws on 1 million Claude conversations and OpenAI’s analysis covered 1.5 million ChatGPT sessions. Anthropic said its published research reflects targeted questions and interests, while OpenAI did not respond to requests for comment.

Anthropic added that it supports external independent research. The AI Observatory researchers cautioned that their dataset, drawn from voluntarily shared conversations, likely underrepresents sensitive uses that people may be reluctant to disclose, so the findings should be treated as a lower bound rather than a full picture.

For Malaysian policymakers, the gaps highlight the risk of relying solely on industry data when drafting rules on generative AI use, content safety and platform accountability.

Related: OpenAI

Reporting based on MIT Technology Review. Figures and claims are subject to revision as the story develops. DomainFork publishes editorial context, not investment advice — see our editorial standards.