Why our AI future needs fact-checkers
The Washington Post (archive.ph)
By Tim Gordon
2023-02-22 18:43:24GMT
Tim Gordon is a trustee for Full Fact, a fact-checking charity based in Britain, and the co-founder of Best Practice AI, an artificial intelligence strategy and governance advisory firm.
New artificial intelligence tools promise to unleash a wave of innovation. Unchecked, these same tools threaten to swamp us with misinformation and falsehoods.
ChatGPT, the most popular of the Generative AI tools, automates creation of human-level writing. It does so by predicting the most likely thing that a human being would have written next in an any given situation. The results — as numerous breathless commentators will attest — are impressive, from credible-sounding academic texts to Shakespearean sonnets generated from just short prompts written in plain English.
But they’re only as good as the material they have to work with. These platforms extrapolate from the huge corpus of text that has been uploaded to the internet over the past few decades. Given what humanity has actually been writing about in those years, this input data can trigger troubling output. This ranges from factually false statements to outright bias against minorities. The belated, and limited, demonstration of Bard, the ChatGPT competitor built by Google’s vaunted AI team, delivered a disputed astronomical claim on its first outing.
Meanwhile, it turns out that algorithms designed to answer questions and sound human will cut corners to meet this prime directive — “hallucinating” or making up their replies if necessary. The same question posed twice can elicit two radically different answers. Both articulated in an equally confident tone.
This mass of misinformation does not even require bad actors. Every child who decides that ChatGPT could do their homework for them or any adult on a deadline risks becoming a vector for infection.
The content generated does not yet match the best human output. But speed, scale and low cost give the algorithms a huge advantage. AI-powered systems producing personalized content at the blink of an eye are well positioned to dominate the content distribution platforms — such as TikTok, Twitter, Google and Facebook — that increasingly mediate our media consumption choices. The bulk of future content on the internet will be produced by AI. In turn, this content will train the next family of generative AI tools.
This is clearly an important risk — and it is a familiar one, similar to the social networks’ battle with online hate. When an algorithm is built to do something — to maximize attention or create pleasing words — then that is what it will default to. This is the original sin that threatens to pollute the fabulous new source of wealth that generative AI represents. Sam Altman, chief executive of OpenAI, the nonprofit turned for-profit company that built ChatGPT, clearly understands the risks that the new systems pose. His focus means that some of the smartest people on the planet, backed by nearly unlimited resources, are working on this. They are in a race: A new field of hacking — prompt injection attacks — dedicated to taking down the programmatic guardrails on ChatGPT has emerged.
If Altman’s teams do not succeed, the effect will be to turn our information world on its head. Fact-checkers currently seek to identify, isolate and remedy outbreaks of “fake news.” In the future, we might need to assume that everything is infected until proved otherwise.
So what might be done? Tools such as ChatGPT need to find their place in the information hierarchy, ideally as a conversational front-end to high-quality information-retrieval systems, as the current alliance with Microsoft’s Bing search engine hopes to do.
One idea is to identify reliably hygienic information sources whose provenance, process and editorial culture can be trusted and even audited. These could provide the training and source data for online fact-checking tools that will need to become as ubiquitous as spell-checking software currently is. But this will not be a panacea — if nothing else, automating a fact check is far more computationally complex than generating a plausible-sounding claim.
Moreover, who decides what is true is always a political issue. It is existential in autocracies and largely left to the market in democracies. Fact-checking organizations can become rallying points for this conversation, bringing voices beyond the commercial and the powerful to the table.
Democratic societies need to up our game, fast. We can, for example, look to Finland, where they proactively educate their citizens on both misinformation and AI. This is seen as a critical component of both domestic security and 21st-century citizenship. AI is getting very good at providing answers — as humans, we must train ourselves to keep asking the right questions.
If we can surf the generative AI wave, then it might yet carry us to a revolution in productivity, education and even the human condition. But waves can be dangerous — and we do not yet know how to swim.
The Washington Post (archive.ph)
By Tim Gordon
2023-02-22 18:43:24GMT
Tim Gordon is a trustee for Full Fact, a fact-checking charity based in Britain, and the co-founder of Best Practice AI, an artificial intelligence strategy and governance advisory firm.
New artificial intelligence tools promise to unleash a wave of innovation. Unchecked, these same tools threaten to swamp us with misinformation and falsehoods.
ChatGPT, the most popular of the Generative AI tools, automates creation of human-level writing. It does so by predicting the most likely thing that a human being would have written next in an any given situation. The results — as numerous breathless commentators will attest — are impressive, from credible-sounding academic texts to Shakespearean sonnets generated from just short prompts written in plain English.
But they’re only as good as the material they have to work with. These platforms extrapolate from the huge corpus of text that has been uploaded to the internet over the past few decades. Given what humanity has actually been writing about in those years, this input data can trigger troubling output. This ranges from factually false statements to outright bias against minorities. The belated, and limited, demonstration of Bard, the ChatGPT competitor built by Google’s vaunted AI team, delivered a disputed astronomical claim on its first outing.
Meanwhile, it turns out that algorithms designed to answer questions and sound human will cut corners to meet this prime directive — “hallucinating” or making up their replies if necessary. The same question posed twice can elicit two radically different answers. Both articulated in an equally confident tone.
This mass of misinformation does not even require bad actors. Every child who decides that ChatGPT could do their homework for them or any adult on a deadline risks becoming a vector for infection.
The content generated does not yet match the best human output. But speed, scale and low cost give the algorithms a huge advantage. AI-powered systems producing personalized content at the blink of an eye are well positioned to dominate the content distribution platforms — such as TikTok, Twitter, Google and Facebook — that increasingly mediate our media consumption choices. The bulk of future content on the internet will be produced by AI. In turn, this content will train the next family of generative AI tools.
This is clearly an important risk — and it is a familiar one, similar to the social networks’ battle with online hate. When an algorithm is built to do something — to maximize attention or create pleasing words — then that is what it will default to. This is the original sin that threatens to pollute the fabulous new source of wealth that generative AI represents. Sam Altman, chief executive of OpenAI, the nonprofit turned for-profit company that built ChatGPT, clearly understands the risks that the new systems pose. His focus means that some of the smartest people on the planet, backed by nearly unlimited resources, are working on this. They are in a race: A new field of hacking — prompt injection attacks — dedicated to taking down the programmatic guardrails on ChatGPT has emerged.
If Altman’s teams do not succeed, the effect will be to turn our information world on its head. Fact-checkers currently seek to identify, isolate and remedy outbreaks of “fake news.” In the future, we might need to assume that everything is infected until proved otherwise.
So what might be done? Tools such as ChatGPT need to find their place in the information hierarchy, ideally as a conversational front-end to high-quality information-retrieval systems, as the current alliance with Microsoft’s Bing search engine hopes to do.
One idea is to identify reliably hygienic information sources whose provenance, process and editorial culture can be trusted and even audited. These could provide the training and source data for online fact-checking tools that will need to become as ubiquitous as spell-checking software currently is. But this will not be a panacea — if nothing else, automating a fact check is far more computationally complex than generating a plausible-sounding claim.
Moreover, who decides what is true is always a political issue. It is existential in autocracies and largely left to the market in democracies. Fact-checking organizations can become rallying points for this conversation, bringing voices beyond the commercial and the powerful to the table.
Democratic societies need to up our game, fast. We can, for example, look to Finland, where they proactively educate their citizens on both misinformation and AI. This is seen as a critical component of both domestic security and 21st-century citizenship. AI is getting very good at providing answers — as humans, we must train ourselves to keep asking the right questions.
If we can surf the generative AI wave, then it might yet carry us to a revolution in productivity, education and even the human condition. But waves can be dangerous — and we do not yet know how to swim.
