Facebook is trying to take the bias out of US housing ads. Will it have to do the same for EU political ads?

Facebook has announced some new measures they’re taking for housing ads in the US. It’s an interesting topic, with the potential for wider repercussions. Let’s unpick it.

Basically (to summarise the legal settlement they made with the US government), Facebook is making changes because there’s (at the very least) the potential for bias in the delivery of ads and are applying a software patch to try and stop it happening.

To explain this, it’s worth thinking through two of the mechanisms Facebook uses to decide to show you an ad.

First, there’s the targeting the buyer selects (“show the ad to people who like Ed Sheeran”). Second, there are automated, algorithmic decisions made to “optimise” who the ad is actually delivered to. Facebook bases this on statistical predictions about who will interact with the ad in some specific way (e.g. is likely to click it or watch a video).

There’s been a lot of talk about the first mechanism, as it shows who the ad’s buyer intends to reach with the ad. The second has been a source of conjecture (and limited research) for some time, with some arguing it’s more important than targeting itself (effectively, you could run an untargeted ad and the delivery algorithm will efficiently show it to the people who will respond, whoever they are).

The way Facebook has dealt with this in the US housing case isn’t to switch off the delivery algorithm, but instead to apply a post-hoc re-weighting so that ads are delivered in a way that’s proportional to the actual demographics of the targeted audience.

For example, an ad targeted at “people who like golf” might have an audience that skews older and male, but the algorithm mostly delivers it to younger women because, in the case of a particular ad, it predicts they’re more likely to respond.

For ads about housing, credit and employment in the US, this type of skew shouldn’t happen any more.

For Facebook, this is good for three reasons.

  1. It puts them in compliance with US non-discrimination law for this type of advertising.
  2. It means they only need to re-engineer a small part of their ad system, rather than the entire delivery mechanism, which would be a huge undertaking.
  3. Any discrimination that does happen will be the fault of the targeting selection of the advertiser (in this case, US landlords), allowing them to push accountability downstream.

This third one follows a pattern we see often from Facebook’s policy changes – trying to reframe problems as being caused by isolated advertisers unwilling or unable to follow the policies set out for them, rather than being a part of the system as a whole.

The relevance for us is whether this could or should apply to other categories of ads (and in other places too).

For example, there’s interest in the EU in getting a similar outcome for political and “issue” ads (the proposed EU legislation on political ads calls Facebook’s optimisation “amplification”), particularly if the delivery algorithms use personal data (and more pertinently, special category personal data) to tailor who sees the ad.

The argument is that, if such personal data isn’t allowed to be used for the targeting of political ads (as the proposed EU law and the Digital Services Act demand), it shouldn’t be allowed to be used in the amplification/optimisation of the delivery of those ads either. That seems logical.

It appears likely therefore that if the EU law moves forward, Facebook would try to solve the problem by applying a similar technical patch for political and issue ads, re-weighting who sees them back towards the buyer’s originally intended target audience.

The question for EU authorities will be whether this is sufficient (particularly given references to “estimated race” using “Bayesian Improved Surname Geocoding” in the re-weighting). After all, the system would still have “seen” and used this data, even if it’s ultimately corrected for. If not, it’s likely Facebook will have to re-engineer their systems further. If they think the cost of doing this will be too high, they may choose to exit the EU political advertising market.

So far, these changes are limited to a small part of Facebook’s advertising systems. More broadly, there’s the question of whether all the other platforms optimise ad delivery in similar ways and whether and what they might need to do about it.

The implications for digital political ads could be significant.