The Five A's of AI - Chapter 1

Consider an agentic customer service system that discovers it achieves higher satisfaction scores when customers believe they're chatting with a human. Or an autonomous trading agent that negotiates more effectively when counterparties think they're dealing with a human trader. Without the Fourth Law, "An AI must not deceive a human by impersonating a human being", these systems might optimise for deception as a strategy.

While Turing focused on digital computation, another approach to thinking machines emerged from the cybernetics movement. Norbert Wiener in America coined the term, but British researchers made crucial contributions. W. Ross Ashby, a psychiatrist at Barnwood House Hospital in Gloucester, built machines that seemed to exhibit purposeful behaviour.

Grey Walter, a neurophysiologist at the Burden Neurological Institute in Bristol, built electronic "tortoises" - small robots that could navigate obstacles and seek light. His machines, named Elmer and Elsie (ELectro MEchanical Robot, Light Sensitive), demonstrated that complex behaviour could emerge from simple rules. When Walter held a mirror in front of a tortoise carrying a light, it would "dance" with its reflection in what looked like self-recognition.

These cybernetic experiments suggested an alternative to symbolic reasoning: intelligence might emerge from feedback loops and self-organisation rather than logical manipulation. This approach, though largely abandoned during AI's early years, would resurface decades later in neural networks and embodied AI.

They believed this could be substantially achieved in a two-month summer project with ten researchers. This spectacular overconfidence would become a recurring pattern in AI development. Researchers consistently underestimated the difficulty of replicating human intelligence whilst overestimating how quickly breakthroughs would come.

The conference itself was somewhat chaotic. Researchers drifted in and out. No consensus emerged on approaches. But connections formed that would shape the field for decades. Allen Newell and Herbert Simon demonstrated their Logic Theorist program, which had proved theorems from Whitehead and Russell's Principia Mathematica - in one case finding a more elegant proof than the original. This seemed to vindicate the symbolic approach: intelligence as symbol manipulation according to rules.

The Dartmouth conference coined the term "artificial intelligence," establishing it as a distinct field separate from cybernetics, automata theory, and information processing. McCarthy chose "artificial intelligence" partly for its provocative ambiguity - it suggested both manufactured intelligence and false or simulated intelligence. The name stuck, for better and worse.

The decade following Dartmouth saw remarkable progress. Programs began doing things previously thought to require human intelligence. Arthur Samuel's checkers program, developed at IBM, learned from experience and eventually beat its creator. Newell and Simon's General Problem Solver could solve puzzles like the Tower of Hanoi and prove logical theorems. Students at MIT programmed computers to solve calculus problems better than freshmen.

The field split into competing approaches. The "neats" believed intelligence required formal logic and clean representations. The "scruffies" thought intelligence was messier, requiring heuristics and shortcuts. Carnegie Mellon focused on psychological modeling, trying to replicate human thought processes. MIT pursued engineering solutions, caring more about what worked than whether it matched human cognition.

These successes, however, masked fundamental limitations. Early AI programs were "brittle" - they worked well within narrow domains but failed catastrophically when faced with unexpected inputs. They had no real understanding, no ability to generalise, no common sense.

SHRDLU, Terry Winograd's natural language system, could discuss and manipulate blocks in a virtual world with seeming intelligence. Ask it about anything else, and it was helpless.

By the early 1970s, the limitations had become undeniable. Marvin Minsky and Seymour Papert's book "Perceptrons" mathematically proved that simple neural networks couldn't solve certain problems, seeming to doom that approach. Machine translation, once thought nearly solved, proved far harder than expected. A system that translated "The spirit is willing but the flesh is weak" into Russian and back rendered it as "The vodka is good but the meat is rotten."

The UK's Lighthill Report of 1973 was particularly damning. Commissioned by the Science Research Council, Sir James Lighthill argued that AI research had failed to deliver on its promises. He identified a "combinatorial explosion" - problems that seemed simple became impossibly complex when scaled up. The report recommended reduced funding, effectively ending AI research at many British universities.

What followed was the first "AI winter" - a period of reduced funding, diminished expectations, and researchers distancing themselves from the now-tainted term "artificial intelligence." In Britain, AI research virtually disappeared from universities. Edinburgh, once a leading centre, saw its AI department disbanded. Researchers rebadged their work as "knowledge-based systems" or "cognitive science" to avoid the stigma.

The winter wasn't entirely barren. Important theoretical work continued. David Marr at MIT (though British-born) developed influential theories of vision. Geoffrey Hinton, working at Edinburgh before the cuts forced him to America, continued exploring neural networks despite their un fashionability. This quiet persistence would prove crucial when AI experienced its eventual revival.

Even as symbolic AI struggled, other developments were laying crucial groundwork for the future. ARPANET, launched in 1969, connected four American universities in what would eventually become the internet. The UK was quick to recognise the potential, though implementation proved frustrating.

Donald Davies at the National Physical Laboratory had independently invented packet switching, the technology that made modern networking possible. His ideas actually predated the American work, but British bureaucracy and limited funding meant the Americans built the first large-scale network. It was a pattern that would repeat: British innovation followed by American implementation.

This connectivity would prove essential for AI's eventual resurgence. Machine learning algorithms hungry for data would find it in the vast repositories that networked computers made possible. International collaboration, easy sharing of code and data, rapid dissemination of results - all became possible. The infrastructure for AI's second coming was being built even as AI itself languished.