This is part II of a series of some practical suggestions to deal with incoming AI change, to preserve society’s well being. I welcome any commentary or further suggestions.
This suggestion is geared toward not immediate disasters involving large power grabs by an AI, but by addressing possible undue influence it could exert on society while being hard to scrutinize.
For example, AI could be paid to influence elections, public knowledge, financial markets, and more to a seeming significant danger.
(II.1) I suggest AI companies not be allowed to receive money intended to influence their datasets, training sets, or conversational output. They are still allowed to place clearly marked ads as they see fit. For example, by using a section with a background of significantly different color and indicating ‘Ad’.
Any company should also not to be allowed to alter its training set or have its training procedure made in a way that strongly promotes the company itself, be it selling other products, encouraging investing in the company, or strongly misrepresenting its own capabilities, trustworthiness, expertise and moral judgement.
This would also apply to Open Source models used by any company. For example, they are not allowed to donate to a 3rd party institute to produce an OS model as they see fit, as they could employ those models to circumvent the law.
(II.2) Moreover, the training set (for example, at least the document title of each document employed), and also an accurate overview of its training method should be publically accessible and searchable as long as the model is commercially available. This includes any Reinforcement Learning methods employed.
Think of this as food labels telling the ingredients of a food item. An AI should have a label of its dataset to ensure it is not manipulative or promotes behavior contrary to our ethics.
Practical AI risk II: Training transparency
This is part II of a series of some practical suggestions to deal with incoming AI change, to preserve society’s well being. I welcome any commentary or further suggestions.
This suggestion is geared toward not immediate disasters involving large power grabs by an AI, but by addressing possible undue influence it could exert on society while being hard to scrutinize.
For example, AI could be paid to influence elections, public knowledge, financial markets, and more to a seeming significant danger.
(II.1) I suggest AI companies not be allowed to receive money intended to influence their datasets, training sets, or conversational output. They are still allowed to place clearly marked ads as they see fit. For example, by using a section with a background of significantly different color and indicating ‘Ad’.
Any company should also not to be allowed to alter its training set or have its training procedure made in a way that strongly promotes the company itself, be it selling other products, encouraging investing in the company, or strongly misrepresenting its own capabilities, trustworthiness, expertise and moral judgement.
This would also apply to Open Source models used by any company. For example, they are not allowed to donate to a 3rd party institute to produce an OS model as they see fit, as they could employ those models to circumvent the law.
(II.2) Moreover, the training set (for example, at least the document title of each document employed), and also an accurate overview of its training method should be publically accessible and searchable as long as the model is commercially available. This includes any Reinforcement Learning methods employed.
Think of this as food labels telling the ingredients of a food item. An AI should have a label of its dataset to ensure it is not manipulative or promotes behavior contrary to our ethics.
Thank you for reading.
See also: Practical AI risk I: Watching large compute