The loss of trust will leave large gaps for intelligent algorithms to step into. Part of the reason we seem to be happy to hand over control is because we are less confident about our place in the world. We may be seeing the beginning of an era where our own knowledge and perspectives become augmented by technology and so we will shift from trusting human knowledge and experience to trusting Wikipedia. If we aren’t kind of there already.
We need to understand the role that AI is likely to play in shaping the way human trust evolves, and therefore how financial institutions, whose core business involves garnering trust, will evolve.
Trust is a funny thing. It feels like a rational assessment but is actually a human emotion that we feel, a kind of calmness about something inherently uncertain. We feel it when we engage with people that are similar to us, when we hear things that confirm our pre-conceptions, and ultimately, we feel it when we have elevated levels of oxytocin in our body. It is a chemical/psychological feature that screens the world to help us understanding what/who we should listen to, and therefore what is actually true. Human trust is the cornerstone of asset management, banking and insurance industries to name a few.
Earlier this year, Morningstar released a report highlighting that passively managed equities had, for the first time, exceeded actively managed. While a useful data-point, it is far from the triumph of algorithms. The growth of passive funds is a combination of intelligent automation and a general disbelief that the typical asset manager can consistently add value above a benchmark after fees. It is a kind of own goal for the industry, a failure of human trust and storytelling, rather than an acknowledgement that algorithms can ‘do it better’. It is the absence of trust, that allowed a more mechanical solution to take hold.
Does science inspire trust?
Firms that use algorithms face difficult problems inspiring trust in non-technical audiences. Algorithmically-based funds are generally very strong in framing uncertainty in a statistical light, and inspiring scientific curiosity and fascination. For technically minded researchers, the world begins as a swirling mass of uncertainty, and slowly starts to take shape as layers of data and models are added. For asset management, the forecasting clarity that is available is limited, and the best models are only slightly better than chance.
Resting on a mountain of mathematical talent, engineers, data scientists and scientific folk often assume that trust is inspired by a statistically better model, rather than something more human. It doesn’t, a fact that quickly emerges when faced with disappointing results. Therefore, quantitative managers use similar storytelling to explain their models, relying on human intuition to inspire familiarity and therefore trust. Connecting the logical strands of inherently non-linear, complex and statistical constructs to a random assortment of heuristic understandings is, at best, a tricky undertaking.
Technology companies have turned to more basic psychological ‘nudges’ to push us to trust and use their products. Automated telephone bots that modulate the frequency of their tone to match ours, building robots that smaller and more frail than us, and placing AI as a ‘helper’ not a competitor are but some of the devices used by AI companies to gain our trust in this technology. But ultimately, they have relied on the notion that technology is our tool, and competitive forces require that we should embrace it. Like any tools however, we will ruthlessly discard them if they don’t serve our needs, and so the ‘trust’ we place in them is utility-based, and is very different to what we would place in a human relationships that are inherently uncertain with allowance to fail.
Like a genie out of the bottle, AI is also playing dangerous game with trust around the world. Take GPT-2 – an artificial intelligence system built by OpenAI that uses artificial intelligence to synthesize new text. Given a short paragraph, GPT-2 can continue the text in the same style. The software generates text that is indistinguishable from that written by a human, and is very difficult to differentiate from major publications. OpenAI has not released the code globally, though it has been replicated. It is a problem for organisations that use bots to write their stories, as without a human author behind it, will become unverifiable. It could also be a problem for financial markets which can move strongly based on sentiment, especially around difficult political topics.
If GPT-2 makes you think twice about what to trust next time you see a news story, Deepfakes will do the same for video. Deepfakes is a technology that essentially ‘moves’ the facial expressions of a person on a video based on a different person’s face, a kind of remote puppet master. While early versions of this technology could be detected with the naked eye, it incorporates adversarial systems, which is constantly improving this capability. In other words, we may be able to falsify visual news stories, interviews, and other material in real-time, from just about anywhere. This makes believing what you see, not only what you read, as difficult.
Finally, the issue of bias in algorithms is a fascinating one, and has garnered great interest from regulators, academics and practitioners alike. All predictive models rely on data, and the shape of that data to make predictions. A certain class of algorithms, those that profile us for loans, insurance, or predict our criminal capability, can carry biases that we may not be entirely comfortable with in our current society. Even if they don’t, the threat that they may do, and that the algorithm may be biasing against us for real or imagined reasons is something that we are and will increasingly contend with.
The scale of the decision making that AI can do for us is already unmatched by human labour, so it is unlikely that we can step back to human judgement. Instead this is more likely to be an exercise in strengthening our trust in algorithms as a society.
With trust in mechanical and algorithmic solutions, will investment management glide toward a more transaction, utility-like industry? There are certainly strong early signs, as the pool of passive money grows, and active management continues to struggle with trust. But we are not yet at a place where the algorithms used in investment management can conjure the emotion of trust to match a good human storyteller. Perhaps with the torrent of human-emulating algorithms, we will be.