India Brings AI Voice Calls and Caller-ID Reports Under Anti-Spam Rules

01India Brings AI Voice Calls Under Anti-Spam Rules and Requires Caller-ID Apps to Send User Reports to Telcos

India has expanded its commercial-communications rules to cover two growing sources of spam intelligence and automated outreach: reports collected by caller-identification apps and calls placed through software or AI voice systems. The Telecom Regulatory Authority of India, or TRAI, said Friday that qualifying apps must send users’ spam reports to telecom operators, while businesses must disclose automatically initiated calls in advance.

The amendments connect caller-ID and call-management apps to a blockchain-based anti-spam platform run by telecom operators. Apps that allow users to flag calls as spam or junk must transmit those reports to the platform, where operators track commercial communications and enforce anti-spam rules. TRAI said the change is intended to enlarge the pool of reports available for action against spammers.

The requirement directly affects services such as Truecaller, a Stockholm-based caller-identification and spam-blocking company that combines community reports with automated detection and other signals. India accounts for more than 350 million of Truecaller’s over 500 million monthly active users worldwide. Truecaller called the mandatory transfer a “one-way exchange” and anti-competitive, arguing that commercially valuable data would move from apps to operators; that is the company’s position, not a demonstrated effect of the rules.

The amendments retain a separate restriction barring call-management apps from automatically blocking, filtering or labeling calls from government-designated number series used for promotional, service and transactional communications. Users may still block individual calls from those ranges on their own devices.

TRAI also brought automatically initiated calls into its application-to-person, or A2P, framework. A2P calls are calls generated without a person directly dialing the recipient and include robocalls, prerecorded messages and calls using artificial voices. Businesses using these systems must declare both the system and the associated phone numbers to their telecom operators before calling. Undeclared A2P calls will be treated as spam.

Operators may charge up to 5 paise per minute to terminate A2P calls, although calls using certain designated number ranges are exempt. The rules regulate disclosure rather than prohibiting businesses from using AI or automated calling technology.

Important implementation details remain unresolved. TRAI has not specified exactly what information apps must submit, how users should be notified or asked to consent, or how long transferred data may be retained and used. It also has not said whether built-in reporting tools in Android, iOS or other phone dialers are covered, or confirmed the final enforcement mechanism. Whether software-assisted calls initiated by a person fall within A2P also remains ambiguous.

02Anthropic Opens Model Evaluations to Accenture Staff as Each Plans at Least $1 Billion Over Five Years

Anthropic is bringing outside evaluators inside its AI lab for the first time, giving Accenture personnel employee-like access to examine how frontier models are trained and deployed. The arrangement puts into practice Anthropic CEO Dario Amodei’s proposal for “embedded evaluation,” intended to provide continuous scrutiny beyond the external tests commonly conducted before a model’s release.

The work will be led by Faculty, the specialist AI business Accenture acquired in January. Its personnel will evaluate and red-team Anthropic’s models, assess whether their behavior aligns with intended goals, and test safeguards. They will also be able to observe models during training, follow decisions governing development and deployment, and speak directly with Anthropic employees.

Anthropic says this access should allow evaluators to check whether the company is meeting its safety commitments, identify blind spots, report incidents, and provide the public with a better-informed account of models’ benefits and risks. Unlike conventional external evaluators brought in around a release, embedded evaluators are meant to watch both the technology and company decisions as they develop.

Anthropic and Accenture each expect to invest at least $1 billion over the next five years to build capacity in this area. The commitments are plans, not amounts already spent, and neither source specifies how much will directly fund evaluation work.

The independence question remains unresolved. Anthropic will pay Accenture directly because pooled and government funding systems do not yet exist. It argues that Accenture’s status as a large public company offers functional independence and says the arrangement makes Anthropic’s accountability more verifiable rather than transferring responsibility. But the company acknowledges there is no settled standard for evaluator access, disclosure, or funding independence.

The partnership is non-exclusive: Anthropic plans to announce additional evaluators in the coming weeks, while Accenture may perform similar work for other AI developers. Anthropic is also discussing self-funded pilot programs with the nonprofit evaluator METR and other organizations.

Exactly who will receive incident reports, what findings will become public, and how disputes will be handled remain unspecified. The parties have also not disclosed the evaluators’ headcount, access boundaries, start date, or public reporting format.

03Independent test says OpenAI ad identifier can report activity across websites and link it to ChatGPT accounts

OpenAI requires businesses advertising in ChatGPT to install tracking code on their websites, much like Meta or Google advertising pixels. An independent technical investigation now claims that code can send browsing activity back to OpenAI with an identifier tied to a visitor’s ChatGPT account.

The investigation examined __obi, a one-year cookie that OpenAI classifies as analytical rather than marketing. Its author says two independent capture methods reproduced the mechanism on a phone, while several months of traffic covering 936 advertiser pixels across 1,029 hostnames provided a broader comparison.

According to the findings, the ChatGPT client generates a random identifier and requests a short-lived, signed token from OpenAI. For a signed-in user, that token contains both the identifier and an account subject. The client then sends the token to OpenAI’s advertising collector, which writes the same identifier into the __obi cookie on the .openai.com domain.

The cookie is configured for cross-site requests. When an advertiser’s page loads OpenAI’s pixel software or submits an event, the browser can attach __obi before the software itself runs. On the author’s device, one identifier reached OpenAI from 12 commercial websites through 13 pixel IDs; every request received a 202 accepted response.

Those requests can include the page’s origin and path, but the investigation found no query strings among 23,929 observed URLs. Reported paths nevertheless exposed activities such as viewing a medical-condition page, entering a debt-solutions funnel or visiting a litigation intake form.

The pixel software can also obtain identity fields supplied by an advertiser or read them from forms, rendered page text and tag-manager data. Email addresses, phone numbers, and first and last names were SHA-256 hashed before transmission—a transformation that does not itself establish anonymity. Country, region, city and postal code were sent in plain text.

OpenAI maintains separate analytics and marketing consent settings. The author says every decoded synchronization token recorded analytics consent, and that allowing analytics while rejecting marketing still triggered __obi. OpenAI has not confirmed that result or explained how it uses, retains or shares the collected data.

OpenAI Support acknowledged the author’s questions and said the observations would be referred for internal review, but did not answer why __obi is classified as analytical or whether users who reject marketing consent still receive it.