AI Policy

Automated systems operate in five broad areas. Ranking and recommendation decides the ordering of your feed, the reels you are shown, and which accounts and topics are suggested to you. Integrity and safety scans content for nudity, graphic violence, self-harm promotion, scams, impersonation, spam networks and coordinated inauthentic behaviour. Support runs iTweet AI, which handles reports, appeals, billing questions and recovery guidance. Creation includes optional tools such as caption suggestions, alt-text generation, translation, background removal and reply drafting. Advertising uses interest signals to decide which promoted content is eligible to be shown to you.

Every one of these systems is a tool, not an authority. Ranking does not decide what is true; it decides what is likely to be relevant. Safety classifiers do not decide guilt; they raise a signal for enforcement, and consequential outcomes involve a human reviewer.

Recommendation systems use your on-platform behaviour: what you view and for how long, what you like, save, share and skip, who you follow and interact with, searches you run, and coarse location or language signals where you have permitted them. Safety classifiers operate on publicly posted content and on content that has been reported to iTweet. Support conversations operate on what you type into iTweet AI plus the account context needed to answer you.

End-to-end encrypted one-to-one messages are not read by any automated system. iTweet does not hold the keys, so message content cannot be scanned, classified, summarised or used for recommendations or training. Where you choose to report a message, only the portion you submit becomes visible for review.

iTweet does not sell your data to model developers. Aggregated and de-identified signals may be used to improve iTweet’s own ranking and safety systems — for example, learning that a class of content consistently produces reports helps the classifier catch it earlier. Where a third-party model provider processes a prompt on iTweet’s behalf, it does so under contractual terms that prohibit using that content to train the provider’s general-purpose models.

Private content — encrypted messages, draft posts you never publish, and material in accounts set to private — is not used to improve public recommendation models. Where personalisation is not wanted, the personalisation controls in settings let you reset interest signals and reduce the depth of behavioural profiling.

Synthetic media is allowed on iTweet, but it must not be passed off as real when the difference matters. Where content is materially generated or altered by AI and depicts a real person saying or doing something they did not, a real event that did not happen, or a scene presented as documentary footage, it must be clearly labelled. Creative, obviously stylised or clearly fictional output does not require a label.

iTweet may add a label itself where automated detection or a report indicates synthetic origin. Repeatedly publishing unlabelled deceptive synthetic media is treated as a serious violation and can lead to loss of reach, loss of monetisation and account termination.

You may not use AI tools, on or off iTweet, to produce sexual imagery of a real person without their consent, to create sexual content involving minors in any form, to impersonate a real person for deception, fraud or harassment, to generate voice clones used for scams, to fabricate evidence of a crime, to manufacture medical or electoral misinformation, or to run networks of automated accounts that simulate organic engagement.

Automated scraping of iTweet content to build datasets or train external models is prohibited without written permission, as is bypassing rate limits or authentication to collect content at scale. Non-consensual intimate synthetic imagery is a zero-tolerance category and results in immediate removal and termination.

Where an automated system produces a consequential outcome — content removal, feature restriction, loss of monetisation eligibility, or account termination — you are entitled to a human review. Appeals are read by a reviewer who was not part of the original decision, and the reviewer can overturn the machine outcome entirely. If your appeal succeeds, the restriction is lifted and the record is cleared.

Submit an appeal through the appeal form. Ranking outcomes — how much reach a post got — are not appealable, because reach is a consequence of audience behaviour rather than an enforcement decision.

iTweet AI can be wrong. It does not provide legal, medical or financial advice, it cannot make binding commitments on iTweet’s behalf, and it will not disclose another user’s personal information. Treat its output as guidance and verify anything consequential. Where a matter requires a human decision — a legal request, a serious safety escalation, a disputed termination — the conversation is routed to a specialist team rather than resolved by the model.

Classifiers are evaluated for uneven performance across languages and regions, since a system tuned on English text often underperforms on Indian language content. Where a measurable gap is found, the affected category is routed to human review more aggressively until the gap closes. Ranking changes are tested on a limited slice of traffic before wider release, and safety-relevant models are re-evaluated on a fixed schedule.

Model inputs are transmitted over encrypted channels, retained only as long as needed to produce and audit the output, and scoped to the minimum context required. Support conversations are retained for a limited period to allow appeals and quality review, then deleted. Access to raw model logs is restricted to a small engineering group and is itself logged.

This policy is reviewed as the systems change. Material changes are announced in the app before taking effect, and aggregate statistics on automated enforcement are published in the Transparency Center.