Can UltoPulto remove me from an AI model?

Training turns a very large collection of examples into a set of numbers, the model weights. An image of you that was in the training data is not stored inside the model as a file. It nudged millions or billions of weights by a tiny amount, together with every other example.

Because that influence is spread across the whole model, there is no record to look up and delete. Researchers study ways to remove the influence of specific data after training, a field called machine unlearning, but exact removal generally means retraining the model without that data, which is slow and costly for large models, and approximate methods are still an area of research.

That is why a promise to "delete you from AI" is not one any company can keep today. It is also why UltoPulto's public materials, product screens and request letters never use that language.

• Source images. If a photo is taken down from the page where it was published, later crawls and later datasets built from fresh crawls will not find it there.
• Datasets and dataset indexes. Many image datasets are lists of links to images rather than the images themselves. Dataset hosts and the sites that publish them can be asked to remove entries.
• AI providers. Companies that train and operate models are organisations that process personal data. Where a privacy law applies to you, they must answer requests for access, deletion and objection, including requests to stop using your data for future training.
• Future collection. Do-not-train registries, site-level signals and content credentials tell crawlers and trainers your preference. They depend on the trainer honouring them.
• Outputs. If a system is producing your personal data or your likeness, that is a separate problem with its own remedies, including provider complaint routes and platform takedown rules.

None of these changes a model that already exists. They reduce what is available for the next one, create a legal record of your objection, and remove the public copies that other systems would otherwise keep finding.

In the EU and UK, the GDPR gives people a right to erasure and a right to object to processing. In December 2024 the European Data Protection Board published an opinion on AI models and personal data, which makes clear that models trained on personal data are not automatically anonymous and that data protection rights and obligations apply to their development and use. See EU privacy rights.

In California, the CCPA gives residents the right to ask a business to delete personal information it collected and to opt out of its sale or sharing. See California privacy rights.

How these rights apply to weights that have already been trained is still being worked out by regulators and courts. A request is still worth making: it must be answered, it puts your objection on record, and it can stop your data being used in future training.

For each AI provider in its registry, UltoPulto prepares an access, deletion and objection request under the law that applies where you live. Where the provider accepts email it is sent as your authorized agent; where the provider only accepts a web form, you get the prepared text and the steps. Each request has a deadline, and replies are logged as evidence.

The request asks the provider to delete personal data from datasets, caches and indexes it controls, to add your identifiers to its suppression list for future training, and, where it says removal from trained weights is not feasible, to say so in writing and describe what it does to prevent your personal data appearing in outputs. See the methodology and the transparency page.