Who to sue when the robot loses your money

The first case when people went to court about the loss of investment, which resulted in the work of autonomous machines, will be a test of the limits of reliability

Every day, robots are becoming more and more like humans, but you still cannot sue them.

Therefore, the Hong Kong magnate chose a different option for the lack of a better one. He is suing the seller who convinced him to entrust part of his capital to a supercomputer, trading on the stock exchange cost him more than $ 20 million.
')
Samatur Lee Kin-kan, whose father is one of the largest investors of Shaftesbury Plc, which owns most of London’s Chinatown, Covent Garden and Carnaby Street, is facing off against Rafael Costa, who has been selling investment funds to such companies for most of his career. as Man Group Plc and GLG Partners Inc. This is the first time that people went to court about the loss of investment, which resulted in the operation of autonomous machines, will be a test of the limits of reliability, and the black box problem brought to light: if people do not know how a computer makes decisions, who is responsible when Will something go wrong?

“People believe that algorithms work faster than people and better make decisions,” said Mark Lemli, a law professor at Stanford University who runs the local law, science and technology program. “Often it is, but when it turns out wrong, or when algorithms go astray, investors look for a scapegoat.”


Rafael Costa

The deadlines that preceded the litigation were taken from documents sent to court on commercial matters in London, where the hearing of the case is due to begin in April of the following year. It all started at lunch in a restaurant in Dubai on March 19, 2017. Then Lee, 45 years old, met Costa, a 49-year-old Italian, known in his environment as "Captain Wizard." During the meal, Costa described a robotic hedge fund, whose services for fully automatic money management using AI will soon begin to be offered by its London-based company Tyndaris Investments.

The K1 supercomputer developed by the Austrian company 42.cx will be combing online information sources such as news articles and social networks to assess investors' mood and make predictions about the future value of futures. Then he will send stock management instructions to the broker, eventually adjusting the strategy based on the information received.

The idea of ​​a fully automatic money manager instantly inspired Lee. He met with Kosta at dinner three days later, and before that in an email he wrote to him that the AI ​​foundation was “exactly what I needed.”

In the following months, Costa shared with Lee the results of simulations in which K1 achieved double-digit returns on invested money, although now they argue about the thoroughness of historical testing. Lee eventually entrusted K1 with managing $ 2.5 billion - of which $ 250 million was his own, and the remainder was leverage from Citigroup Inc. This amount was planned to double.

However, Lee's fascination with K1 immediately passed when the computer started trading at the end of 2017. By February 2018, he steadily lost money, including the case when $ 20 million was lost in one day - February 14 - because of the stop loss ["moose" traders jargon / approx. transl.], which, according to Lee’s lawyers, would not have been enforced if K1 were as sophisticated as Costa claimed.

Lee now sued Tyndaris for $ 23 million due to the exaggeration of the supercomputer's capabilities. Lawyers for Tyndaris, who sued Li for $ 3 of unpaid fines, deny that Costa exaggerated the possibilities of K1. They claim that he never guaranteed that the AI ​​strategy would be profitable.

Sarah Makatomaini, a lawyer at investment firm Lee, who is suing Tyndaris, declined to comment. Rob White, a spokesman for Tyndaris, refused to link the editorial board with Costa for an interview.

Legal battles are the forerunner of what can happen when AI is included in all walks of life, from work mobiles to virtual assistants. When technology misfires, the question of who to blame for this is controversial. In March, US prosecutors acquitted Uber Technologies Inc., who was accused of the death of a 49-year-old pedestrian who died under the wheels of one of her robotic vehicles.


Profit in the 2019th year for every $ 100 of investments invested in 2014; The profit of the S & P 500 is based on re-invested dividends.

In the world of hedge funds, attempts to use AI have become a necessity after many years of unsatisfactory work of people-managers. Quantitative investors - computers specifically designed for finding and selling trades - are already popular. Less common are pure AI funds, automatically trained and improved performance based on experience, with no specially programmed data. As soon as the AI ​​develops its own understanding of the situation, even its creators no longer understand why he makes certain decisions.

“You may find yourself in such a situation that you cannot explain why you are holding this position,” said Anthony Todd, co-founder of Aspect Capital, a London-based company experimenting with AI strategies, before allowing them to work with client investments. "One of our concerns about using machine learning techniques is the loss of clear hypotheses about the market behavior."

Lee's lawyers claim that Costa won confidence by inflating the qualifications of technicians who wrote the K1 algorithm, and said, for example, that they were developing Deep Blue, an IBM chess computer that marked the dawn of the AI ​​era when he beat the world champion in 1997- m


Garry Kasparov plays against IBM Deep Blue in 1997

In a conversation with Bloomberg, 42.cx founder Daniel Matts said that not one of the informatics specialists who consulted him was related to Deep Blue, but one of them, Vladimir Lvovich Arlazarov , in the 1960s developed the chess program “Kaissa "[She won the 1st computer world championship in Stockholm in 1974 / approx. trans.]. He acknowledged that such experience may not be particularly strongly associated with investments. Algorithms have learned quite well how to win games from people, because they have clear rules that can be simulated - which is not true of stock markets. Alazarov told Bloomberg that he gave general advice to Matts, but did not work on K1.

Inspired by a 2015 European Central Bank study that explored investor sentiment based on Twitter, 42.cx created software that could generate mood signals, said Matts, who recently agreed to pay $ 17 million to the US Securities Commission to settle a lawsuit about cheating investors of their mobile payments company Jumio Inc., committed in the early 2010s. And how to respond to these signals was decided in Tyndaris, he said.

“It was a great program,” Matts told us over the phone. - Signals issued by us had clear scientific substantiations. I think we worked pretty well. I know that I can capture moods. I am not a trader.

The court documents contain a lot of controversy over whether Li was deceived about the possibilities of K1. For example, the car created a single request in the morning if it received a clear signal about the mood of investors, although Lee says that he believed that she would trade during the day. Costa’s lawyers say he told Lee that buying or selling futures based on several trading signals were part of the company's plans, but were not planned from the very beginning.

There were periods of several days when K1 did not trade at all, because it did not find sufficiently strong trends. In one of the messages to Kostya Lee, he complained that K1 is resting, while quietly passing adverse events, in the hope that it will not reach the activation of the stop-loss order (a pre-set level, upon reaching which the broker will sell shares to limit losses during a sudden fall prices).

This happened on the day of St. Valentine in 2018 In the morning, K1 placed an order with the broker Goldman Sachs Group Inc. At $ 1.5 million in futures for the S & P 500, predicting a rising index. But he went in the opposite direction, when it turned out that inflation in the US rose faster than expected, which led to activation of the stop loss K1 by 1.4% and the loss of $ 20.5 million by the fund. However, after a few hours, S & P regained its position - and this, according to Lee’s lawyers, shows that the stop loss limit for K1 was set “rude and wrong”.

Lee states that he was assured that K1 would use its own “deep learning opportunities” daily to determine the correct stop loss limit based on market factors such as volatility. Costa denies saying such a thing, and declares that he said that people will always appoint this level.

In an interview, Matts said that K1 did not lay the possibility of setting a stop loss on his own - he only knows how to give two types of mood signals: a generic that Tyndaris can use to enter a position, and a dynamic one that can be used to exit or change a position. Although Tyndaris advertised a fund managed by K1 to other investors, the spokesman refused to answer the question of whether he managed any real money. And all references to the supercomputer disappeared from his site last month.

Investors, and, in particular, Markus Storr, say they are worried about entering AI funds on the stock exchange, especially considering that funds that used AI as their main strategy earned less than half of the S & P 500 profit in three years, according to indices from Eurekahedge AI Hedge Fund.

“We cannot judge the code,” said Storr, who makes decisions about investing in Bad Homburg, a German trust fund. “So we can only judge the research capabilities and structure.”

But what happens when companies use chatbots to sell products to customers? Perhaps it’s not even possible to sue the seller, added Karishma Paroha, a London lawyer at Kennedys, specializing in product liability. “Misunderstanding refers to what the person tells you,” she said. “What happens when someone sells something to you?”