AI Inquiry Atlas

An unofficial, independent guide to the public records of the Joint Select Committee on Artificial Intelligence. Not connected with the Parliament of Australia or the committee. Its makers also made a submission (no. 481; see About). Summaries and scores are written by AI and can be wrong: check the original before relying on them.

Submission 357

Name withheld

Submitter group: Individuals.

Position labels are not shown for individual submitters.

An AI model identified the submitter group and chose the extracts. Counts do not measure importance or merit.

Open in the interactive guide

Extracts from the submission, in the submitter’s own words.

Extracts

Quotes are the submitter’s own words. Text taken from documents, including scanned pages read by text-recognition software, can differ from the PDF. Check the original. Listed in the same order as on the submission’s page in the interactive guide.

  1. The person should be shown a separate question which is easy to understand. The answer should be set to “no” unless the person actively chooses “yes”. In other words: No clear opt-in means no AI training.

  2. It should also not be able to make the service much worse, remove unrelated features, or repeatedly ask the same question until the person gives in. If refusing consent comes with a major penalty, then the consent is not really being given freely.

  3. However, the law can allow some secondary uses without consent when they are related to the original purpose and reasonably expected by the person. This does not give Australians a simple and certain rule about AI training.

  4. That does not automatically mean that the person has agreed to have the information collected in bulk, kept indefinitely, analysed for sensitive information, or used to train a commercial AI model. Public visibility and consent to AI training are not the same thing.

  5. This is a difficult technical problem. However, it should not mean that a company is automatically allowed to keep using a valuable model which was built using information it had no right to use.

  6. Information accidentally collected from people who never agreed should not be used to train AI.

Extracts from recommendations

Extracts chosen by an AI model, in the submitter’s own words. Listed in the same order as on its page in the interactive guide.

  1. require a clear and separate opt-in before personal data collected for another purpose can be used to train, fine-tune, develop or improve AI;

  2. The person should be shown a separate question which is easy to understand. The answer should be set to “no” unless the person actively chooses “yes”.

  3. Real consent should be clear, informed and freely given. This means that: the question should be separate from general terms and conditions; the company should explain what kinds of personal data it wants to use; the company should explain why it wants to use the data; the person should have to actively choose “yes”; silence or inactivity should not count as consent; a box which is already selected should not count as consent; the company should make “no” just as easy to select as “yes”; and the person should be able to change their mind later.

  4. Companies should also not be allowed to pressure people into agreeing. For example, a company should not be able to offer the normal service for $25 per month if a person agrees to AI training, but charge $100 per month if the person refuses. It should also not be able to make the service much worse, remove unrelated features, or repeatedly ask the same question until the person gives in.

  5. The law should therefore cover: conversations and prompts; documents and uploaded files; private code; images and video; voice recordings and written transcripts of recordings; location and movement information; health information; feedback and relevant account information; and new information which an AI system infers or creates about an identifiable person.

  6. A company should not avoid the rule by calling AI training “product improvement”, “quality assurance” or “research”. The law should look at what the company is actually doing with the information.

  7. The rule should also apply when information is obtained from a data broker, contractor, related company, customer or public website. A company should have to check where the information came from and whether it had permission to use it.

  8. Relevant exemptions for small businesses, employee records and political organisations should not create gaps.

  9. Companies should have to keep records which show: where the information came from; why it was originally collected; whether the person opted in; what the company was allowed to use it for; how long it could be kept; which other organisations received it; and which training datasets and AI models used it. Regulators should be able to examine these records and check whether information from people who did not opt in was kept out of the training process.

  10. require independent checks and impact assessments for high-risk AI projects;

  11. Projects involving large amounts of personal data, children, health data, biometric information or important decisions should also require a privacy and human rights impact assessment before training begins.

  12. give people clear rights to information, withdrawal, correction and complaint;

  13. People should have the right to ask a company what categories of their personal data it collected and whether those categories were made available for AI training. The company should explain where the information came from, what it was used for, and which other companies received it. However, the law should not require a company to make an impossible or misleading claim about the exact effect one piece of information had on billions of numbers inside a trained AI model. The company should instead have to provide the information it can genuinely verify from its records.

  14. A person should be able to withdraw consent as easily as they gave it. Once consent is withdrawn, the company should immediately stop using that person’s information for future AI training. Where possible, the company should also remove the information from active source datasets, prepared training datasets, indexes and caches. It should tell contractors and other relevant organisations that they must also act. Withdrawing consent should not result in punishment or the loss of unrelated services.

  15. give regulators the power to deal with datasets and models created through unlawful practices;

  16. Depending on how serious the breach was, a regulator or court should be able to order the company to: stop using the information for future training; delete the information from source and prepared datasets; retrain the model; use reasonable machine-unlearning or model-editing methods; prevent the model from producing the affected information; restrict or stop use of the model; pay for an independent expert to check what has been done; or destroy the model in a serious case.

  17. Devices should clearly show when they are listening, recording or sending information. People should have simple physical controls and should be able to see and delete recordings and transcripts.

  18. Information accidentally collected from people who never agreed should not be used to train AI.

  19. A company should still have to collect only what it needs, keep the information secure, delete it when it is no longer needed, and avoid unfair or discriminatory uses.

  20. Some types of AI surveillance may be too intrusive to permit through a consent box. The European Union has prohibited untargeted scraping of facial images for facial-recognition databases, certain uses of biometric information to infer sensitive characteristics, and most emotion recognition in workplaces and schools. Australia should consider similar limits.

  21. create stronger protections for children, workers, vulnerable people, bystanders, biometric information and Indigenous data; and

  22. Children, employees, people with disabilities, and people who depend on essential services may also have less freedom to refuse. They need stronger protection than an ordinary consent form can provide.

  23. Aboriginal and Torres Strait Islander languages, stories, images and cultural knowledge may belong to or affect a community as well as an individual. Indigenous communities should have a genuine role in deciding how that information is used, supported by Indigenous data sovereignty and fair benefit-sharing.

  24. ensure that Australian rules also apply to overseas companies doing business in Australia.

  25. prevent companies from using unreasonable prices, service restrictions or confusing designs to pressure people into agreeing;

Read the original

Open in the interactive guide