Thread by @divyasiddarth

'How AI Fails Us', out with @SafraCenter today! ethics.harvard.edu/how-ai-fails-us

I am really proud of this work! What started as a conversation with @glenweyl has grown into an incredible collaboration, with @DrDaronAcemoglu @dsallentess @katecrawford @profjamesevans & Michael Jordan. 1/

We may differ in our prescriptions, but we share a core belief: That the current direction of AI is both unproductive and dangerous. And that better futures, for technology and for humanity, are possible. But we have to act now. 2/

Venture capitalists and the largest technology companies are investing billions of dollars into a vision of AI characterized by autonomous machine intelligence that aims to achieve and surpass a generality ascribed to human intelligence. 3/

They assert that these investments will lead to broadly beneficial futures for humanity. These are not credible promises. 4/

First, the target of replicating, with the aim of surpassing, human-level cognitive capabilities in autonomous systems sets the relationship between humans and machines as one of competition, rather than one of cooperation and augmentation. 5/

This excessively displaces workers, mechanizes our conception of human capacities, and forgoes myriad opportunities for improving human productivity. Simply put, this isn't the best use of technological resources to support human flourishing - not by a long shot. 6/

Second, the focus on achieving a singular and autonomous 'general intelligence' necessitates a drive towards concentrating resources, compute, data, and investment into an ever-shrinking set of organizations and people. Human competition + Autonomy = Centralization. 7/

This level of extreme concentration in the direction of productive resources is not a new idea. Which means that history is littered with failed, often disastrous attempts. 8/

There is no reason to believe this new iteration will fare any better than those that came before - with results ranging from miscalculation-driven famine to unlivable cities to collapsed post-Soviet economies. 9/

The lesson is clear: to contend with the real crises facing the planet, we should not invest vast resources in small, centralized groups pursuing extremely narrow goals. 10/

So, then what? Is the answer to give up on the possibility of technological progress? No. If there is one thing I hope readers take from this piece, it is this: there are new dreams, better dreams and plenty. 11/

Debate on these points is often framed as “safety/ethics” v. “power/technology/capabilities”. This is a false dichotomy. Actually existing AI forgoes power and capability in its narrow focus: replacing it with a pluralist paradigm is both the productive and the ethical path. 12/

(We don’t need to look far: the history of digital technology is full of alternatives. Norman Weiner’s ‘The Human Use of Human Beings’ offered alternatives to the Turing Test and narrow conceptions of intelligence before McCarthy had even coined the term AI.) 13/

And the scope of technological pluralism becomes truly appreciable when we read the work of historians who have worked to recover technical visions from communities and contexts outside institutional conversations (A Brock, M Hicks, S Dick). 14/

Today's ecosystem is similarly full of alternative directions. Briefly: work that develops systems unrelated to human intelligence (ex. AlphaFold), mechanism design / social tech approaches to rework preference aggregation and participation (ex. pol.is), 15/

Work to build systems that collaborate with humans (ex Cobots), interaction paradigms that incorporate human expertise (ex. machine teaching), systems that build precision through personalized data (data coops), visions of an 'AI services' model to replace general AI. 16/

Most exciting to me are projects that break entirely with the AI trajectory, and build new paradigms for decentralized agency: DAOs, data trusts, knowledge commons, platform coops, wikis, collaborative design projects, P2P standards + protocols, digital democratic platforms. 17/

These are diverse directions and the landscape is constantly shifting. What is clear is that a future of digital pluralism is possible: one that builds on this work, and views intelligence as an emergent outgrowth of human and nonhuman systems interacting across scales. 18/

While some in the AI space consider superintelligence a natural evolutionary step from human intelligence, we contend that this ecological, interdependent, and complex interplay of intelligent systems is far closer to true evolutionary progress. 19/

This work builds on the critical contributions of so many brilliant folks calling for and building a better vision for the future of technology. We hope to work with all of you to create this future together. From the singularity, to the plurality! 20/