Algorithm auditing...not here yet, but on its way
They wield incalculable power, but where’s the accountability? The time for independent verification of algorithms is only around the corner - despite the headaches it will cause, writes Dominic Mills
A search for the term ‘algorithm auditing’ doesn’t throw up much in the way of depth.
But that’s hardly surprising. Although our lives are increasingly ruled by algorithms, as a society we are only just beginning to think through some of the consequences of this. From insurance, to schools, to job hunting, to prison sentencing, they are everywhere.
Oh yeah, and then there’s the few billion in £/$ of advertising money whose deployment and potency hangs on the algorithm.
It’s astonishing if you think about it. All this economic and societal activity is based on a giant, blind, leap of faith. Paradoxically, the very black-box nature of the algorithm underscores our perceptions of its infallibility. It’s Little Britain’s ‘the computer says yes/no’ taken to its ultimate conclusion.
Algorithm efficacy to one side, there are ethical and moral dimensions to consider also. They have incalculable power, but where’s the accountability?
But pull the piece of string hard enough, as some are beginning to do, and it becomes clear: at some point there needs to be some independent verification of the algorithms deployed both corporates and public bodies - let’s call this an audit.
The movement for algorithm auditing, if it can be called such a thing, is led by Cathy O’Neil, who describes herself as an algorithm activist and is the author of Weapons of Math Destruction.
I confess neither algorithms or auditing are subjects that consume many of my waking moments - let alone putting the two together. This is a niche, niche world.
My interest was sparked by an almost-random comment last week by a leading media consultant. I put the issue to a handful of individuals embedded, in different ways, in our business. The collective reaction can be summed up in four ways (and nothing from one individual who has forgotten to turn their out-of-office off two weeks after they returned - get with the action, pal.)
1. Wow, that is an interesting concept.
2. No, I don’t know of any instances of algorithm auditing in our business. I’ve never seen it asked for and I've never seen it written into a contract.
3. Nobody will allow their algorithms to be audited.
4. I can see where it has value.
As yet, however, it has yet to touch our world. Indeed, the difficulties advertisers have had mandating even standard audits of their media activity suggests getting to algorithm auditing will be like pushing water uphill.
But ultimately algorithm auditing is about risk control and evaluation, and that is a subject never far from the corporate boardroom table. From there, it may not be a big jump to media.
There is action on the supply side - the likes of Deloitte, PwC and Accenture are clearly feeling their way into the subject, as this Deloitte paper makes clear. For the big four auditors and other consultancies - and I don’t mean this in a cynical way - algorithm auditing is potentially a lucrative business line, so it is in their interests to push it.
The demand will lag but, you could argue, a reputational, brand safety or performance failure on the part of a major advertiser may prompt them to start asking tougher questions either of their agency or the publisher. I don’t know how they have got away with it so far, but when platforms transgress a brand’s safety guidelines, blaming it all on the algorithm does not seem long-term tenable.
The question then is how you apply algorithm auditing in the media world. At one end of the spectrum, you might have to follow them all the way through the long and complex supply chain - DSPs, SSPs, exchanges, all those myriad suppliers, and all the way to the publisher. This might, for example, be auditing to understand the robustness of verification or viewability measurement.
But does anyone really want to look under the bonnet in that much detail? And who would allow it?
Certainly the big platforms will defend theirs from outside scrutiny to their last breath. They will claim, and you can’t argue with it, that it is their secret sauce. In any case, they are under no financial pressure to do so. Only legislative pressure could make them cede ground, although in the current climate that remains a possibility.
Other publishers will be less able to resist, but would it be worth it long-term for them? It might be if, say, they were able to use a kitemark scheme.
Then there is the question of what are you trying to audit. How specific is the decisioning under scrutiny and what are the parameters? Is it realistic and measurable?
It might be if, say, an advertiser deployed money against an algorithmic model that used proxies, rather than actuals, to predict or assign behaviour. You would be entitled to know, wouldn’t you, the basis of the modelling. Certainly that is one line a consultant with experience of the auditing market thinks: to assess attribution models.
One argument is that performance is all that matters and you either get the performance you wanted, or were promised, or you didn’t. Less need for an audit then.
So, to be clear, at the moment we are talking theory and possibility. Demand currently lags, although that may be chicken and egg. Once the possibility is in place, the demand may follow.
But there is a sense that the winds are blowing in the right direction. First, there is the gradual realisation that we have given our lives over to powerful tools without any independent assessment of whether they do what they claim or how they do it. It’s a subject rising up the agenda. Then there are increasing advertiser demands for accountability, of which auditing is a subset, and algorithm auditing a sub-subset.
If (or when) the two come together in a meaningful way, then we might be into a whole new game. And these things have a way of creeping up on us faster than we realise.