As you will know (from that annoyingly chirpy person striding across our TV screens), over the last few years OTA aggregator Trivago has run extensive TV advertising, claiming, among other things, that Trivago can help you find the “best price” for a hotel room.
Australia’s consumer protection regulator, the ACCC, litigated these claims.
The Federal Court of Australia last week delivered judgement in favour of the ACCC.
The judgment found that that Trivago’s claims were misleading or deceptive given the way that the Trivago algorithms worked, and that Trivago had accordingly breached the Australian Consumer Law by making the claims.
Damages are yet to be assessed.
Although a judgement at first instance, the case is in my view quite novel, and very interesting for us (would-be or real) geeks, for reasons that follow.
Trivago used (and uses) a complex AI algorithm to determine which hotel rooms to show a consumer for a given search. The main thrust of the ACCC allegations was that this algorithm was not predominantly or reliably driven by the price of the room, but rather (in a significant number of cases) by the value of the bid made by the listing entity to Trivago for listing of that particular offer.
Although on the facts the case against Trivago looked relatively straightforward for the ACCC to argue, the more difficult tasks for the ACCC were:
(1) Proving to the judge that the rankings were misleading as to ‘best price’.
The ACCC led expert evidence of a data scientist using test data to infer how Trivago’s underlying algorithms appeared to weight whether and how much a hotel or onine booking entity paid Trivago by way of CPC bid fee and thereby influence, and sometimes determine, the rank or listing of a hotel.
Trivago led conflicting data science expert evidence.
The two experts (Victor Bajanov from Quantium, briefed by the ACCC, and Professor Parkes from Harvard, briefed by the Trivago), after hundreds of pages of analysis, published four different versions of the relevant answers.
(2) Proving to the judge how the rank of listings might affect the decision making by a browsing consumer, by expert evidence of a behavioural economist as to application of heuristics by many consumers.
The judgement is particularly interesting because it addresses two relatively novel areas:
(1) Judicial consideration of complex and significantly conflicting data science evidence that inferred the statistical likelihood as to how underlying algorithms (demonstrated from the black box, testing its operation on test data) operated across a diverse set of factors and fact scenarios. Although the experts had confidential access to the algorithm itself and accordingly could see inside the black box, the key evidence relied upon by the judge, informed by the expert evidence, were results
of running test queries, and not deconstruction of the underlying algorithms.
(2) A judge considering how much weight to give to behavioural economics and whether the developing discipline of behavioural economics is now sufficiently settled to be addressed through evidence as to consumer behavioural factors given by an academic economist as an expert assisting the court.
The decision turned on findings that the Trivago website did not quickly and easily identify the cheapest rates available for a hotel room responding to a consumer’s search: the Trivago website did not display offers unless the Online Booking Site’s (i.e. a hotel’s) CPC bid exceeds a minimum threshold set by Trivago.
Accordingly, in at least some cases, the cheapest offer for the hotel room did not appear on the Trivago website.
Further, the expert evidence established that the offer that was given most prominence on the website (being the so-called Top Position Offer) was in many cases not the cheapest offer for the hotel room.
Based on the data they examined, the computer science experts agreed that higher priced offers were selected as the Top Position Offer over alternative lower priced offers in 66.8% of listings.
However, the effect of the CPC bid by the hotel on ranking as the subject of disagreement between the experts: even on the evidence of the expert briefed by Trivago, the CPC bid was a very significant factor in determining the Top Position Offer – it was (on the evidence of the expert briefed by Trivago) the second most important factor, with a relative importance of between 33.8% and 44.8%.
As algorithms become increasingly more central for how consumers and businesses interact, the decision should be a cautionary tale for any business that makes claims about how the business makes recommendations or determines rankings where those claims are inconsistent with how its algorithms actually work across the broad range of possible search or other use scenarios. This isn’t a straightforward issue, and it can’t be settled with a short description of the algorithm: a superficial or helicopter view assessment won’t be sufficient to convince a judge.
Also, many organisations today have deployed complex algorithms at scale, but may not fully understand the emergent outcomes that these algorithms create – that is, the eventual outputs and their key drivers when the algorithm is deployed in the wild, which may be different to what the algorithms looks like on paper. And if an algorithm is non-compliant for even a relatively small number of use cases, this may still create legal exposure.
The full judgment can be read using the link below.
Peter Leonard is a data, content and technology business consultant and lawyer advising data-driven business and government agencies. Peter is principal of Data Synergies and a Professor of Practice at UNSW Business School (IT Systems and Management, and Business and Taxation Law). Peter chairs the IoTAA’s Data Access, Use and Privacy work stream, the Law Society of New South Wales’ Privacy and Data Committee and the Australian Computer Society’s AI Ethics Technical Committee. He serves on a number of corporate and advisory boards, including of the NSW Data Analytics Centre.
Editor’s note: this piece was updated on 3/2/2020 with some minor clarifications
