AI: Investment, M&A and Liability
The Rise of AI Corporate Activity “Artificial intelligence is one of the most exciting and transformative opportunities of our time.” – Nathan Benaich, Playfair Capital Nicola… Read more
The Rise of AI Corporate Activity
“Artificial intelligence is one of the most exciting and transformative opportunities of our time.” – Nathan Benaich, Playfair Capital
Nicola Blackwood MP, Chair of the UK Government Science and Technology Committee, recently announced an inquiry into the opportunities of robotics and artificial intelligence for the UK. “It is important that the UK is ready with the research, innovation and skills to be able to fully take advantage of the opportunities and manage any risks,” she said, as “the global market for the AI sector is expected to grow to $2-6 trillion by 2025.”
Certainly, artificial intelligence has become something of a hot sector of late. The largest software businesses in the world continue to invest big in AI, whether by way of development such as IBM’s $1bn commitment to Watson or Google’s run of acquisitions including DeepMind Technologies (bought for $625m in 2014 and whose AlphaGo system recently defeated the world champion Go player), Dark Blue Labs and Vision Factory. Apple has been busy buying up Emotient (emotion-detection technology) and Vocal IQ (speech-processing). In addition, Amazon recently acquired Orbeus, the Silicon Valley startup which specialises in image recognition, and Salesforce has made two major acquisitions already in 2016: MetaMind (AI-based personalisation and customer support solutions) and Prediction IQ (open-source machine learning server). The graphic below is a good visual representation of the major acquisitions in the sector since 2013.
As well as large acquisitions, there are a proliferation of startups looking to leverage machine learning tools. Artificial intelligence platforms are starting to become mainstream in the software industry with developments in systems which can analyse both structured and unstructured data. The impact of AI on existing technology and employment is forecast to be profound. A recent Deloitte report estimated that as many as 36% of jobs in the UK were at risk of being automated.
Yet, despite the talk of AI being the next paradigm shift for the internet, it should be remembered that the AI sector is currently still an immature market. According to Bloomberg data, venture capital funding within the software industry as a whole increased 37% in 2015 to a record $24.5 billion. Altogether, AI investments accounted for a little over $2.75 billion (11% of software investment, and less than 5% of total VC funding in 2015). That’s more than double the figure for 2013, but AI is still trailing behind sectors such as adtech, mobile and business intelligence software. The vast majority of deals done in 2015 were also of comparatively small amounts and overwhelmingly into US businesses. 80% of the VC investments were sub-$5 million in size and 90% of the cash flowed into the US.
As a comparator, the table below shows key UK investments from 2014-2016.
However, despite AI being in its infancy, it is already throwing up interesting challenges for lawyers when advising potential investors or acquirers on the risk profile of artificial intelligence and machine learning platforms.Source Beahurst
AI and Acquisition Risk
The general split of risk on any M&A transaction is that (subject to agreed limitations of liability) losses suffered by the buyer caused by matters arising pre-closing fall to the seller, and those caused by matters arising post-closing are for the buyer. By giving warranties to the buyer and providing statements and supporting documents in disclosure, the seller delivers a snap-shot of the business to the buyer. If the disclosure process has been undertaken properly, the buyer should have a good picture of the value of the target to enable it to agree the price based on its knowledge of the business as at closing. If any statement made by the seller about the company later proves not to be true, and the value of the business acquired is reduced because of it, the buyer has a remedy to claim for breach of warranty and look to recover the loss.
The amount of loss which can be recovered under a breach of warranty claim follows the usual breach of contract principles, including principles of legal causation and remoteness of damage. The principle of legal causation is somewhat ephemeral, but essentially holds that it would be unfair to hold a defendant responsible for certain consequences of a breach in circumstances where there is a break in the “chain of causation” between the breach and the harm suffered. To determine whether the chain of causation has been broken by an intervening event, the courts will take into account the foreseeability or inevitability of that event.
Remoteness is a separate principle, but one which also looks to assess whether the consequences caused by the breach should be the responsibility of the defendant (essentially, whether the loss was foreseeable). The test for foreseeability operates in an M&A context such that, subject to any pre-agreed limitations or restrictions in the sale agreement, the buyer can look to recover all losses which are:
- “in the contemplation of the parties” (ie foreseeable);
- foreseeable not merely as being possible, but as being “not unlikely”; and
- foreseeable at the date of the sale agreement, not the date of the breach.
The precise circumstances which unfold following the breach do not have to be foreseeable, but the nature of the loss has to be.
To determine whether something was “in the contemplation of the parties”, the courts will firstly consider what occurs “in the ordinary course of things” (common sense) which the parties will have deemed to contemplate, whether they actually did or not. Secondly, the court will also take into account any actual knowledge of the parties themselves of special circumstances outside the ordinary course. Recent case law has recast this in terms of an assessment by the courts of the “presumed intentions” and “common expectation” of the parties as to the assumption of responsibility by the defendant under the contract.
In terms of a warranty claim on a share sale, questions of foreseeability are relatively straightforward (although actually quantifying the loss may be more difficult). The buyer can protect itself from much of the acquisition risk just by ensuring the warranty schedule is drafted appropriately. Let’s say that the seller warrants to the buyer that the target company owns all intellectual property used in the business, but fails to disclose that key software is actually licensed from a third party. If this is subsequently discovered, it would be difficult for the seller to argue that this loss is not foreseeable and that it shouldn’t be responsible for the diminution in value caused by the company not owning the relevant IP.
Even the risk of how a target company’s IT system will operate post-closing can be mitigated by a combination of the right warranties and some technical due diligence on the target’s back-up systems, use of anti-malware or anti-virus protection or how it utilises open source software. However, the acquisition of an artificial intelligence business may require a different approach.
The creativity of AI
The key feature of an AI system which distinguishes it from previous technology is its ability to act autonomously. It may do so within the confines of a definable structure, but many businesses developing AI see this “creativity” as one of the main benefits of artificial intelligence and machine learning.
In his paper Regulating Artificial Intelligence Systems: Risks, Challenges Competencies and Strategies, US attorney Matthew Scherer gives the example of C-Path, a machine learning algorithm for cancer pathology. Researchers suspected that studying the stroma – the supportive tissue surrounding cancer cells – in conjunction with the cancer cells themselves, would help cancer prognosis. C-Path found that the stroma indicator in and of itself gave rise to a more accurate prognosis; ignoring the actual cancer cells was a counter-intuitive conclusion which went against current scientific consensus, but led to a breakthrough.
C-Path’s ability to “think outside the box” is not down to a creative process, but rather the computing speed and brawn to weigh up thousands more data points than a human could successfully analyse without the restriction of perceived general wisdom. Discussing a computer chess program, Nate Silver in The Signal and the Noise: Why So Many Predictions Fail – But Some Don’t, writes:
“We should probably not describe the computer as ‘creative’ for finding the moves; instead, it did so more through the brute force of its calculation speed. But is also had another advantage: it did not let its hang-ups about the right way to play chess get in the way of identifying the right move in those particular circumstances. For a human player, this would have required the creativity and confidence to see beyond the conventional thinking.”
The NUDT Tianhe-2 supercomputer can process 33.86 quadrillion (that’s 33,860 trillion or 33,860,000,000,000,000) floating-point operations per second. Whilst most home computers don’t operate in this range, their computational power allows them to search through many more possibilities than a human in any given time frame (and this power will only increase over time). This allows AI systems to find solutions humans may not have even considered, let alone attempted to put in place. For some developers, this is the entire point of artificial intelligence. In his book On Intelligence, Jeff Hawkins, the inventor of the PalmPilot and Treo smartphone, writes:
“Personally, I am less interested in the obvious applications of intelligent machines. To me, the true benefit and excitement of a new technology is in finding uses for it that were inconceivable before. In what ways will intelligent machines surprise us, and what fantastical capabilities will emerge over time?”
AI designers are designing their platforms precisely so that the algorithms produce unexpected solutions; but if things go wrong and damage is caused, is that foreseeable? The behaviour of a learning AI system depends not only on its initial programming, but its post-design experiences and the data it interacts with once it has been sent out into the world. Let’s say a buyer acquires an AI platform which has been designed to be semi-autonomous and allowed to alter its objectives based on subsequent experiences. If the AI then does something unexpected which causes the buyer loss, should the unforeseeable behaviour break the “chain of causation”? Should the seller be liable even if the system’s designers had no way of knowing that such an event would happen? Should the buyer take responsibility as it knew it was buying an AI system which had the potential to be “creative”? What if the algorithm which gave rise to the unexpected behaviour was open source?
It is market practice in the UK on a company acquisition for the buyer to be prevented from bringing a breach of warranty claim (other than in respect of tax) after a period of around 18-24 months following completion of the deal. The rationale for this is that an 18-24 period should give the buyer sufficient time to conduct a financial and technical audit; it also gives a reasonable amount of time for any third party claims against the target company for actions pre-closing to come out of the woodwork. Self-learning AI systems are different; no amount of human technical analysis will detect all outcomes that an AI system could deliver (otherwise it wouldn’t be a particularly valuable AI system). In that context, allowing the buyer 18-24 months in which to bring a warranty claim seems totally arbitrary.
To date, as we’ve seen, there have been few UK to UK exits for AI businesses as the sector is still young. Yet, as investment continues to flood into AI companies, it won’t be long before we lawyers may be having to wrestle with some vexing questions around how we structure liability for AI behaviour.