51吃瓜

It was the year of generative AI. The twelve-month period between January and December 2022 gave us DALL-E, Midjourney and ChatGPT, powerful tools that put the combined power of a search engine, Wikipedia and a top-notch content generator at our fingertips.

Tools like Bard, Adobe Firefly and Bing AI quickly followed, rapidly expanding the abilities of your average internet user beyond anything we could鈥檝e imagined just a few years ago. With a couple of simple keystrokes, we can now generate captivating images or pages of written content that, this time last year, would鈥檝e taken hours, days, or weeks to produce鈥攅ven for illustrators or writers with years of training.

Indeed, generative AI is changing the landscape beneath our feet鈥攚hile we鈥檙e standing on it. But this pace of innovation comes with risks; namely, of losing our footing and letting algorithms override human discernment. As a recent in the Harvard Business Review highlighted, the creation of fake news and so-called deep fakes poses a major challenge for businesses鈥攁nd even entire countries鈥攊n 2023 and beyond.

Fortunately, innovation in AI is not just producing results for content generation. It鈥檚 also a tool that, when coupled with good, old-fashioned human instinct, can be used to resolve problems in the systems themselves. But before examining these strategies in more detail, it鈥檚 important we understand the real-world threats posed by AI-generated misinformation.

Recognizing the threats

The potential threats of AI-generated content are many, from reputational damage to political manipulation.

I recently in The Guardian that the journal鈥檚 editors received inquiries from readers about articles that were not showing up in its online archives. These were articles that reporters themselves couldn鈥檛 even recall writing. It turns out, they were never written at all. ChatGPT, when prompted by users for information on particular topics, referenced Guardian articles in its output that were completely made up.

If errors or oversights baked into AI models themselves weren鈥檛 concerning enough, there鈥檚 also the possibility of intentional misuse to contend with. A recent identified several risk factors of generative AI use by humans ahead of the 2024 US presidential election. The report raised the specter of convincing yet illegitimate campaign emails, texts, or videos, all generated by AI, which could in turn mislead voters or sow political conflict.

But the threats posed by generative AI aren鈥檛 only big-picture. Potential problems could spring up right on your doorstep. Organizations that overly and uncritically rely on generative AI to meet content production needs could unwittingly be spreading misinformation and causing damage to their reputations. 

Generative AI models are trained on vast amounts of data, and data can be outdated. Data can be incomplete. Data can even be flat-out wrong: generative AI models have shown a marked tendency to 鈥渉allucinate鈥� in these scenarios鈥攖hat is, confidently assert a falsehood as true.

Since the data and information that AI models train on are typically created by humans, who have their own limitations and biases, AI output can be correspondingly limited and biased. In this sense, AI trained on outdated attitudes and perceptions could perpetuate , especially when presented as objective fact鈥攁s AI-generated content so often is.

AI vs. AI

Fortunately, organizations that use generative AI are not prisoners to these risks. There are a number of tools at their disposal to identify and mitigate issues of bad information in AI-generated content. And one of the best tools for this is AI itself.

These processes can even be fun. One method in particular, known as 鈥渁dversarial training,鈥� essentially gamifies fact-checking by pitting two AI models against each other in a contest of wits. During this process, one model is trained to generate content, while the second model is trained to analyze that content for accuracy, flagging anything erroneous. The second model鈥檚 fact-checking reports are then fed back into the first, which corrects its output based on those findings.

We can even juice the power of these fact-checker models by integrating them with third-party sources of knowledge鈥攖he Oxford English Dictionary, Encyclopedia Britannica, newspapers of record or university libraries. These adversarial training systems have developed sophisticated-enough palates to differentiate between fact, fiction and hyperbole.

Here鈥檚 where it gets interesting: The first model, or the 鈥済enerative鈥� model, learns to outsmart the fact-checker, or 鈥渄iscriminative鈥� model, by producing content that is increasingly difficult for the discriminative model to flag as wrong. The result? Steadily more accurate and reliable generative AI outputs over time.

Adding a human element

Although AI can be used to fact-check itself, this doesn鈥檛 make the process hands-off for all humans involved. Far from it. A layer of human review not only ensures delivery of accurate, complete and up-to-date information, it can actually make generative AI systems better at what they do. Just as it tries to outsmart its discriminative nemesis, a generative model can learn from human corrections to improve future results.

What鈥檚 more, internal strategies like this can then be shared between organizations to establish industry-wide standards and even a set of ethics for generative AI use. Organizations should further collaborate with other stakeholders, too鈥攊ncluding researchers, industry experts and policymakers鈥攖o share insights, research findings and best practices.

One such best practice involves data collection efforts that prioritize quality and diversity. This involves careful selection and verification of data sources, by human experts, before they鈥檙e fed into models, taking into consideration not just real-time accuracy, but representativeness, historical context and relevance.

All of us with stakes in making better generative AI products should likewise commit to promoting transparency industry-wide. AI systems are increasingly used in critical fields, like health care, finance and even the justice system. When AI models are involved in decisions that impact peoples鈥� real lives, it鈥檚 essential that all stakeholders understand how such a decision was made and how to spot inconsistencies or inaccuracies that could have major consequences.

There could be consequences of misuse or ethical breaches for the AI user too. A New York lawyer in hot water earlier this year after filing a ChatGPT-generated brief in court that reportedly cited no fewer than six totally made-up cases. He now faces possible sanctions and could lose his law license altogether.

Generative AI modelers therefore shouldn鈥檛 be afraid of sharing documentation on system architecture, data sources and training methodologies, where appropriate. The competition to create the best generative AI models is fierce, to be sure, but we can all benefit from standards that promote better, more reliable, and safer products. The stakes are simply too high to be playing our cards so close to our chests.

The strides taken by generative AI in the last year are only a taste of what鈥檚 to come. We鈥檝e already seen remarkable transformation not just in terms of what models are capable of, but in how humans are using them. And as these changes continue, it鈥檚 critical that our human instinct evolves right along with them. Because AI can only achieve its potential in combination with human oversight, creativity and collaboration.

The views expressed in this article are the author鈥檚 own and do not necessarily reflect 51吃瓜鈥檚 editorial policy.