But that’s taken agency away from users (we don’t control what we see) and has instead left us with conversations full of hate and discord, as well as a growing epidemic of mental-health problems among teens. That’s a far cry from the global, democratized one-world conversation the idealists dreamed of 15 years ago. With many users left adrift and losing faith in these platforms, it’s clear that maximizing revenue has ironically hurt business interests.
Now, with AI starting to make social media much more toxic, platforms and regulators need to act quickly to regain user trust and safeguard our democracy. Here I suggest six technical approaches that platforms should double down on to protect their users. Regulations and laws will play a crucial role in incentivizing or mandating many of these actions. And while these reforms won’t overcome all the problems of mis- and disinformation, they can help stem the tide ahead of elections next year.
1. Verify human users. We need to distinguish humans using social media from bots, holding both accountable if laws or policies are violated. This doesn’t mean divulging identities. Think of how we feel safe enough to hop into a stranger’s car because we see user reviews and know that Uber has verified the driver’s identity. Similarly, social media companies need to authenticate the human behind each account and present reputation-based functionality to inspire accounts to earn trust from the community.
2. Know every source. Knowing the provenance of the content and the time it entered the network can improve trust and safety. As a first step, using a time stamp and an encrypted (and not removable) IP address would ensure an identifiable point of origin. Bad actors and their feeds—discoverable through the chain of custody—could be deprioritized or banned instead of being algorithmically amplified. While VPN traffic may deter detection, platforms can step up efforts to improve identification of VPNs.
3. acknowledge deepfakes. In line with President Biden’s sweeping executive order on AI, which requires the Department of Commerce to evolve guidance for watermarking AI-generated content, platforms should advance evolve detection and labeling tools. One way for platforms to start is to scan an existing database of images and tell the user if an image has no history (Google Images, for example, has begun to do this). AI systems can also be trained to detect the signatures of deepfakes, using large sets of truthful images contrasted with images labeled as fake. Such software can tell you when an image has a high likelihood of being a deepfake, similar to the “spam risk” notice you get on your phone when calls come in from certain numbers.
4. Filter advertisers. Companies can share a “safe list” of advertisers across platforms, approving those who comply with applicable advertising laws and conform professionally to the platforms’ advertising standards. Platforms also need to ramp up their scrutiny of political ads, adding prominent disclaimers if synthetic content is used. Meta, for example, announced this month that it would necessitate political ads to disclose whether they used AI.
5. Use real humans to help. There will, of course, be mistakes, and some untrustworthy content will slip through the protections. But the case of Wikipedia shows that misinformation can be policed by humans who follow clear and highly detailed content rules. Social media companies, too, should publish quality rules for content and enforce them by advance equipping their trust and safety teams, and potentially augmenting those teams by providing tools to volunteers. How humans fend off an avalanche of AI-generated material from chatbots remains to be seen, but the task will be less daunting if trained AI systems are deployed to detect and filter out such content.
