An industry-wide agreement is needed to identify AI-generated content

With the UK general election coming on July 4th, there is very real concern that this election and others around the world are being influenced by artificial intelligence-powered (AI) content generators.

In May, Margaret Beckett, chair of the Joint Committee on the National Security Strategy, said she was concerned to see the huge disparity in approaches and attitudes to managing harmful digital content, based on written evidence submitted to the committee’s inquiry on defending democracy.

Resham Kotecha, head of policy at Open Data Institute warns that the ability to generate realistic deep fakes at scale, and the abilities of conversational generative AI have changed the potential reach and depth of the challenges we now face. Kotecha believes governments should require the disclosure – by political candidates and political campaigns – of the use of AIs and algorithmic systems, so people know whether they are the subject of targeting and if they have been algorithmically selected and information pushed their way.

Technology firms should be looking at the approaches they can take to ensure published information can be identified and classified as either human- or AI-generated,

Tim Callan, chief experience officer at an identity management firm, Sectigo, says governments should prioritise the requirement of built-in encrypted time stamps on content, which can act as a watermark at the time of capture. This, he says, offers a viable way of identifying manipulated content.

But the tech sector needs to do more. Computer Weekly recently spoke to Meta’s vice president of AI research, Joelle Pineau, During the conversation she spoke about why the company did not put a model it had developed for realistic voice synthesis into the public domain as an open model. The model works with a snippet of a human voice recording and is able to generate speech output that sounds the same. Her rationale is that Meta has yet to develop technology that can accurately distinguish between a human’s voice and the synthetic voice.

Given the problem with deep fakes, society will need to prepare for voice synthesis models. Pineau’s approach is an AI model that is effectively based on and trained with the same data as the voice synthesis model, which would be able to somehow ascertain that a voice recording was AI-generated. Wouldn’t It be far easier if everyone agreed to watermark AI-generated content?

Clearly, a dark web of AI models will emerge. But these should be easier for us all to identify. Just as signed apps and antivirus tools help limit the spread of rogue software, analogous technology could evolve in AI.

Computer Weekly believes the tech sector must collaborate on a global digital watermarking standard, embedded deep in the DNA of AI generated content.