Mediaboss Marketing

As I head to the UKs AI Safety Summit Fringe reading the Governments AI discussion paper, some diverse forces have been shaping my own reflections on AI safety: the Gaza-Israel crisis, humanitarian law and principles, and science fiction film, The Creator.

Recently, many of us working across humanitarian aid agencies have been responding to events in the Middle East. As a community, we are calling loudly for parties to de-escalate violence, allow unfettered humanitarian access, and protect civilians.

Ill leave others to write about Israel and Gaza, but the political, legal and emotional complexity of this protracted crisis also lends insight into some of AIs thorniest issues and the human dynamics at their centre.

If we think of AI as the architecture of a new societal order that could perpetuate harmful power structures unless they are actively designed out of the plans, then it feels logical to infer lessons from human contexts where decision-making, bias, neutrality, reliable information, and humanity itself are stress-tested.

Whose safety?

Safety in its simplest definition, means protection from harm. Laws can contribute to safe environments by setting boundaries ranging from national legislation protecting the right to life to international laws of war. Laws exist to keep people safe, they set norms and expectations, but the simple existence of a law does not guarantee it wont be violated.

More fundamentally, safety requires changing behaviour. Safety needs people to share a mindset, values and beliefs to act, or not act, in a way that harms others. Actively practising humanity, compassion, and addressing structural inequality is as critical to individuals safety as is the rule of law.

This applies to AI safety measures just as it does to peace agreements: for either to be meaningful, the process to get there has to be inclusive. If its only elites at the table, if different groups arent invited to share perspectives and shape decisions, they can soon collapse. The benefits and harms AI can confer or perpetuate diverge significantly according to the individual just as someones experience of a humanitarian catastrophe differs depending on their gender, age, ability and so on.

This is why we must build meaningful bridges between staff in frontier AI companies and Global South civil society, as I argue in How to Do Good and Avoid Doing Harm: 7 Actions for Big Tech, Governments and Civil Society. Otherwise, how can AI architects understand the lived experiences of people very different to themselves, and know what safety means to them?

Can humanity be programmed into machines? (Film spoiler alert)

At the end of a day working on Gaza-Israel, I went to see The Creator for what was intended as two distracting hours of light relief. In it, humans go to destructive war with artificial intelligence. (In retrospect, yes, it would have been wise to read the synopsis in advance).

The film opens with the bad guys artificial intelligence having blown up New York City and occupying New Asia. There, AI civilians are embodied in humanlike simulant form living alongside New Asian humans, while AI soldiers resemble a post-pubescent WALL-E.

Meanwhile, the West is fighting for humankinds freedom led by the good guys in the U.S. army.

So far, so similar to every human vs AI terminator movie.

But all is not as it seems, and this is where the film gets interesting. As our assumptions and biases are tested, the narrative raises questions about god-like AI, compassion, the fog of war, and what it means to be human.

For me, the film explores a key question: if it is our humanity that makes us human (which supposedly differentiates us and makes us superior to intelligent machines), then what does it mean when we lose our humanity?

Filmed in South East Asia, The Creator recalls films like Apocalypse Now and the barbarity of the Vietnam war. In one coastal scene, humans attack from the air with overwhelming force. We watch as the U.S. Army locks a missile target onto an individual AI robot on the ground. It runs, but cant escape the lethal blue circle tracking it. It appears scared and runs for cover where a terrified human family are already sheltering.

As they scream in fear realising they will be caught by the blast, the AI stops, reverses with reassuring raised hands, then sprints back towards open ground and immediate death. Recalling Isaac Asimovs Third Law of Robotics, the AIs instinct (or, adapted learning) is to prioritise the humans survival over its own.

Some techno-optimists argue that an AI-led world would be more compassionate than our human-ruled one. Its a polarising debate like that over existential risk and whether doomer arguments distract from existing and near-term harms.

In a weak moment, having never seriously engaged with the concept that an AI-led world could be more compassionate than our own, current world events caused me to run the thought experiment. Could an AI system be programmed to be more compassionate than a human? What is the nature of compassion?

Programming values into machines

As I consider this, I become trapped in an infinite loop. Why? Because its humans who build AI on foundations of human-selected data. Humans are exceptionally complicated and the decisions we make are based on deeply embedded beliefs, world-views and power dynamics. And selected data on which AI is trained can reflect, over-index, or miss that.

As AI systems are woven ever more into the global social fabric, becoming increasingly deferred to for signifiant decisions over citizens lives, will it ever be possible to solve this Escherian puzzle?

As we know both from the culture wars and violent conflict, people regularly fight over which truths and human rights are right. This even influences interpretations of globally agreed rules like the laws of war.

In other TMR articles, Ive discussed the seductive certainty provided by an algorithmic recommendation. And I dont mean the times Spotify suggests Old MacDonald over Miles Davis because a small child has borrowed your phone.

Im talking about harmful decisions being made in an earthquake response in poor areas where fewer people, especially women, have a mobile phone leading to inaccurate (biased) geospatial data. Or someone from the wrong part of a city being denied parole because biased data flags their characteristics as indicators of reoffending risk.

The Chief Commercial Officer of a frontier AI company recently told me, The thing is, you can programme bias out of an AI system. You cant do that with humans.

But as the creators of AI, its still humans deciding (or noticing) what constitutes bias, what that looks like in data and AI, and what is the right way to correct for and mitigate it. Programming bias AI out of AI system includes decisions on which data to include or exclude, how data is labelled, inclusive approaches to generating questions and gathering data (to know what data even exists), and, the small task of tackling the root causes and symptoms of inequality and marginalisation across society.

In short: who will programme the bias out of the human tasked with programming the bias out of the AI system?

The quality of humanity

Plotting a course towards safe AI means first navigating the messiness of humans, of competing human rights, and structural inequalities. AI safety will remain rhetorical unless there is inclusive analysis and informed protection across genders, age, ethnicities, financial status, geography and more.

Of the four principles that traditionally guide humanitarian aid work humanity, independence, impartiality and neutrality humanity is the one which raises fewest debates. As the UK welcomes the leaders of nations and AI companies to iconic Bletchley Park, not every attendee will share the same interpretation of safety.

But as a starting principle, maybe they can coalesce around one: humanity.

Hit Follow above to be alerted to new articles from The Machine Race blog. Share your comments, corrections and suggestions here, on LinkedIn, or on Twitter @TheMachineRace. Listen to Suzy Madigan talking about AI and society on the Standard Issue podcast and see the About page for authors biography. Thanks for reading.

Originally posted here:
AI Safety: Can we programme humanity? | by The Machine Race by ... - Medium