Will Generative AI do more harm than good?
Gary Marcus, who debated Keith Teare at an Open to Debate event in early March 2023 on whether ChatGPT would do more harm than good, says he now returns to that question almost daily and that the negative side of the ledger has grown heavy. In his assessment, the harms are concrete and accumulating: nonconsensual deepfake pornography, more phishing scams, large-scale disinformation, what he calls cognitive surrender, democratic risk from governments subtly shaping large language models, and generative AI used as a surveillance tool. He also cites IP theft against artists and writers and possible broader damage to employment, while noting that job losses so far have been more modest than some predicted.
Much of his current concern centers on security. He points to a flood of hacks that some have labeled "Black October," which Paul Kedrosky has called The Hackening. Marc Benioff, Salesforce's CEO, described the situation as a "Code red," amplifying a bleak post from Anthropic's Jason Clinton. Marcus agrees with both characterizations, while noting that Clinton expects this bad stretch to last only a year or two. That time horizon is the main point of disagreement worth tracking: if the window really is that short, the question becomes whether defenses catch up faster than attackers refine their methods. This is my own reading, but it suggests the security argument is less about whether damage is happening than about how long the current advantage lasts.
On the positive side, Marcus asks whether the risks are worth taking. He cites Torsten Slok, chief economist at Apollo Global Management and at times fairly positive about the technology, as having made a blunt case that many of the positive effects people were led to expect simply are not materializing. Marcus links to the full argument without reproducing its data. He concedes that he has friends who get a lot out of generative AI, mainly for coding, but concludes that at this point it is not hard to argue the technology has on the whole been a net detriment to society. He adds that there is still a chance it could pull down the entire economy.
For developers and product teams, the practical reading is that the debate has shifted from hypothetical risk to operational cost. Security teams are already treating AI-assisted attacks as a distinct threat category, and the commentary Marcus gathers from Benioff and Clinton frames the current period as an acute phase rather than steady state. A likely trade-off is that organizations leaning on generative AI for productivity gains, especially in coding, will be asked to justify those gains against measurable risk: fraud, IP exposure and incident volume. Whether the balance tips back depends on facts not yet settled, including how long the attack surge lasts and whether the productivity evidence Slok questions improves. Marcus says he will return to Open to Debate on Tuesday in Baltimore, partnered with Cathy O'Neil; there will be no public livestream, but edited versions are expected later on YouTube and on NPR affiliates including WNYC.
Why it matters: Teams shipping AI features face rising scrutiny over abuse, IP and security exposure at the same time that promised productivity gains are being questioned publicly. If the current attack surge is as severe as these commentators suggest, security and trust work becomes a first-order product requirement rather than a compliance afterthought—an inference from the stated concerns, not a reported finding. The near-term test is whether organizations can show concrete benefits that outweigh the documented harms.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Things aren’t looking great.