AI recruitment software is 'automated pseudoscience', Cambridge study finds Claims that AI-powered recruitment software can boost diversity of new hires at a workplace were debunked in a study published this week. Advocates of machine learning algorithms trained to analyze body language and predict the emotional intelligence of candidates believe the software provides a fairer way to assess workers if …
"Mackereth also explained these tools are likely trained to look for attributes associated with previous successful candidates, and are, therefore, more likely to recruit similar-looking people instead of promoting diversity."
So, basically, just like all too many human recruiters.
Yep. I've lived through at least three versions of psychometric testing, various personal statements, "technical" tasks (..if this sentence has less than one aardvarks in it then write I'm wasting my time here ____) and other crap. The people you remember from work are the ones that surprised you - in both a good and a bad way - and it's hard to design a recruitment process that finds those surprising people.
Personal recommendations from people you trust is the best way, in my experience, and interviewing the recommender about the candidate is more important than interviewing the candidate.
> it's hard to design a recruitment process
This ^. It's hard, if not borderline impossible, to know how a candidate will turn out after some time, after routine and fatigue have erased the last shreds of motivation he might have had initially, not to mention during periods of stress and increased workload. So recruiters actually rely on completely unrelated but visible standards, like "does that person look/sound like upper class upbringing?" or even "would I date this person?"...
And totally ridiculous pseudo-scientific tests ("Sometimes I hear voices in my head - True/False")...
The biggest problem is that you only get feedback for the candidates that you do select. Most candidates being interviewed will be at least 'ok' were you to employ them. You might think a recruitment process is good (and train your AI on that data) because all the people you recruit are good. In reality it may be that all the candidates you rejected would actually have been excellent but you will never know that.
It's also hard to reconcile the views of the company between hiring and firing time.
When hiring, the company needs role requirement, people requirement and job description documents which will be used to screen and select candidates.
When firing, the main criterion seems to be that they didn't work long-enough hours.
Yes I own headphones. I literally hear voices in my head every day. I know that's probably not what the test question meant, but it is literally what it says.
I feel uncomfortable relying on a company that is run by people who didn't notice that or care about it to send my paychecks in a timely manner. As such this assessment is over, your company failed, and I will now consider other opportunities, potentially including your direct competitors.
Best of luck, and remember, every interview is a two way exchange.
At the height of the psychometric testing fad our Chief Exec decided to have it done to all the directors. Including my then boss. She was so dispirited by the result that she basically gave up on her career and started a family instead. Which is a shame (for me), because she was by some distance the best manager I ever had.
-A.
If you are flooded with applications for a position and have neither the personpower nor the interest to go through them, don't bother with AI tools.
There is a simpler method that is fairer and as good as the AI tool: pick x applications randomly and invite them in for an interview.
It is quick, fair and cheap.
Of course, it also means that you aren't doing your job properly but when you use an AI tool, you aren't doing your job properly either.
When one of your randohires turns out to be "That guy"
Course that goes both ways, as if you interview them and they pull a Jekyll and Hyde on you, you hired them.
That said, if not 100% reliable, a 20 min structured conversation will shake out a LOT of red flags. That and checking up on their factuals will generally get the task done.
If a turkey slips by, suck it up and can them. Cheaper in at will employment states, but still better than dragging the dead weight in the long run. People like that tend to multiply quickly if left to settle in.
> trained to look for attributes associated with previous successful candidates, and are, therefore, more likely to recruit similar-looking people
So assess the applicants without looking at pictures of them!
Instead have the AI analyse what they write. How they filled in application forms. Ask them to write a paragraph about a non-personal (so they don't "leak" personal information) topic.
"Ask them to write a paragraph about a non-personal (so they don't "leak" personal information) topic"
How about actually talking to them about the prospective job? It's almost impossible these days to get through on the phone to a recruiter (at least here in the UK) and when you do they hardly ever seem to actually listen. I guess the primary motivation is least effort to land the commission.- hence the use of "AI".
But this is not a new problem. Way back in the early '90s at a seminar by the CEO of a leading agency I asked whether any client asked them to select for individual excellence. He instantly responded that nobody did. They just want median bodies that "fit in" - essentially meaning they keep their heads down and don't make waves. "AI" (a statistical matching engine really) seems pretty good at finding the median, so maybe it's exactly what's wanted for recruitment.
I worked with a very good group of programmers in Boston back in the 70's, they were writing code in Pascal and added comments to describe the data and processing - all the comments describing the functionality were written in Bahasa Indonesia - since they were writing in Pascal we all worked well together.
Well, that's better than using pictures, but better than actively harmful is a low bar. All the AI will tell you (if it's properly calibrated) from a writing sample is whether you appear to have proper grammar from whatever dataset it has. It's likely to get that wrong anyway. If it was trained on only formal writing, an informal paragraph would probably be judged poorly. It could become overly focused on spelling and mark someone down for using a correctly spelled word that it doesn't know.
Also, it's not any use. The only time you need a writing sample is if you're hiring this person as a writer or in a job where there will be a lot of writing, and depending on the job, you need different writing samples. If you're hiring a journalist to write articles, a sample of an email won't tell you if they can do it. If you want a creative writer of fiction, an open source project's readme won't tell you about their skills. If you're hiring a person to write documentation, being able to understand and explain the function of a system may be more important than writing quality, because correctness and structure in the documentation is often more important than literary style.
This issue is as much about the training data and historical inputs. The way these systems evolve in the process of training intrinsically amplifies bias. So while I agree image processing adds nothing and actively makes things worse, language models are just as bad. Even purely text based ones can't reliably identify bias, are known to magnify it, and the languages people actually write in also add an intractable computing problem on top of everything else. You never get past a weak correlation, you hit the wall of diminishing returns before you get a usable system, and no matter how many times you double your "whatevers", you never get past the asymptotic bound.
You literally cannot succeed in accurately solving this class of problem with the ML models we have today. It is incomputable. The fact that most HR people are statistically marginal at it should have been a red flag, but that never stops the snake oil salesmen.
'Dubbed the "Personality Machine", the system looks for the "big five" personality tropes: extroversion, agreeableness, openness, conscientiousness, and neuroticism'
Just like the polygraph, launched a century back, this makes the fundamental error of inferring subjectively defined behaviours or psychological characteristics from completely uncorrelated physical phenomena. We should have learned by now that it doesn't work (although in the case of the polygraph we obviously haven't).
But of course the reality is not about being right - it's about not being held responsible for decisions that affect people's lives "the machine made the decision, so don't argue with me".
It's also about the (pseudo) scientific veneer: Take the "personality tests" which were all the rage several decades ago: Initially they were real tests used to triage mental health patients in institutions. Obviously using them to screen job applicants is pointless and would at best help eliminate the totally naive (or stupid) psychopaths, those who would indeed not lie to questions like "Sometimes I feel like eating human flesh" or some such... But science, man, science can't be wrong, even (apparently especially) when used out of context.
There is a major problem with your post. You are (I'll be kind) misunderstanding the word "science". NO matter how many times they pretend it is science some areas of knowledge just are not! Example of non-science are: sociology, economics, psychology, climateology etc.
This is especially true when you are told "the science is settled"
"You got an ology!"
Ah, who remembers that advert! And if you are old enough to remember that advert actually being on TV then the AI will probably automatically reject you for being too old.
Any job that requires a personality test is an instant 'no thanks'. They are utter junk.
3 out of 4.
Quite good, but try extending your reading around actual science and, perhaps, reduce time spent on social media science-substitutes.
Alternatively, stick with the social media and you could become an 'influencer'.
I look forward to seeing even better results or a change of direction next term.
All of those can be sciences. All of those and more can be done wrong, thus not scientifically. Science requires rigorous testing of hypotheses until there's some confirmation that they're likely true, and then more testing to make sure that wasn't wrong. That's equally true in chemistry or psychology. Psychology has had more people recently who decided they knew what they were talking about because they made up some theories and stated them as proven, doing it wrong. Chemistry may therefore seem harder, but there was a time when people did the same thing there, for example deciding that heat and fire were created by an invisible substance called phlogiston based on the principle that "we don't know how, and this could be possible, so therefore it is true". That was no better.
There are people in all fields who approach it that way, but you appear to dislike the entire field based on people failing to approach it scientifically. There are others who do it correctly, and you can verify whether a certain study was tested scientifically or not and reproduce the behavior if it was. Either that or you have a preexisting prejudice against some theory, but being unwilling or unable to disprove that theory using the scientific method, you are applying an incorrect stereotype to the field as a whole.
So, if the fantastic employee you hired wore a red t-shirt during their interview, the software will then recommend candidates who also wear red t-shirts. There's really little need to regulate stuff like this, because any company stupid enough to use it will fail for very obvious reasons. But that won't stop some companies from using them : perhaps it will take another "AI" being used to examine company bankruptcies to make the connection between using this software and going bust to do that. It's what happens when you're willing to believe that software can replace the need to use actual intelligence when performing fundamental tasks that are crucial to your business.
And this is a copy of our dress code...
Erm... Do I have to wear red?
Yes, yes. That's the uniform specification for your job title...
OK, but why red?
Because it hides the blood stains.
Blood stains???!!!!
Howard Sway wrote: "There's really little need to regulate stuff like this"
Indeed, I wondered why "The researchers believe the technology needs to be regulated more strictly." Wouldn't it be sufficient to point out it doesn't work and companies should spend their money elsewhere?
I would hope that would convince them, but from the stories others have told about other pseudoscience used in recruiting, I think increased regulation is justified here. Tests that ask questions and claim to produce useful information about a candidate are likely useless, but they're not directly as harmful. AI software that looks at a person's face and environment have already demonstrated on multiple occasions and in most if not all variations, that they are extremely prone to creating artificial stereotypes. Far from diminishing the affect of bias on recruitment, these are likely to accentuate discrimination even if the humans involved would not. The result is no longer a thing of chance, but now actively harming certain groups, which is often illegal under existing laws. That's a valid reason to regulate it under those laws.
The people pushing this are trying to profit of fraudulent claims and information. Ignoring them doesn't just allow them to predate naive companies, it also causes direct harm to applicants who are unfairly rejected.
Ignoring quacks and frauds amounts to enabling them. More active measures are required to keep the problem in check unfortunately.
Never was and never will be based on the existing systems. It's not intelligent in any meaningful sense of the word. Generally or otherwise.
2) Bias in a ML or AI system has no need to be based on human anything, but human biases will be quite hard to remove from human built systems or human managed data. But if an AI could be built, it wouldn't need human biases to make it over that bar. Cognition is not an intrinsically human trait.
3) Your last point is a racist dog-whistle. If you were not aware of that please stop. If you are, please sod off.
Tripped at the finish line, number 3. Requiring diversity is racism, and there is no other word for it. Race, and gender for that matter, should not be a consideration at all. Saying a complaint about diversity is a "racist dog whistle" is only saying "that's a different kind of racism than the racism I support." Just because your favored form of racism is favored by certain groups does not make it any more acceptable than if it's favored by other groups. If you didn't understand this now you know, please stop. If you did then sod off yourself.
If you want true hiring equality, assign all candidates a number and have them add their resume leaving off all personal information, including dates, and only submit the nonpersonalized resumes to hiring managers. They won't know what they're hiring until after the job offer is accepted by the candidate, and there will be no bias that way.
I think I mentioned it before, but what got me my last job against an equally qualified guy was my response to a question about Flemming's left and right hand rules. I correctly stated them but then added I could never remember which was which and would have to look them up.
I stayed with that firm for nearly 20 years - then retired at 70.
I hadn't heard of or thought about Flemming's rules since 2nd secondary school year science lessons (now Year 8 I think) back in 89-90.
I suppose it's good testament to my science teacher that I remembered what they were and what the fingers were meant to represent (but would still need to look up which finger or thumb was what)
On the other hand I still resent / laugh at his report about me for that year in his class... it was short.. a whole year of classes summed up in 4 words.
"This Child Is Dangerous"
I still don't know what I did to warrant that report, but at least it gave my parents a laugh.
"Competence" is something that's really only possible to validate after the fact. We'll find out if they were really a good hire during their probation period.
By that point however the "Talent Acquisition Team" have already done their bit, got their signing bonus, and moved on.
How well someone fits the company "Leadership and Diversity Principles" (Did they go to school with one of our directors? Are they from a minority group we don't have a lot of at the moment?) these are things that they can check off a list before we hire someone, so that's what they care about.
Hmm... "Talent Acquisition Team" by its name acquires talent to the house. When the Acquisition Team was formed, there was no Talent Acquisition Team present, thus no talented people could be acquired to the Team. This leads me to conclusion that the Team consist of untalented people, and therefore cannot really acquire talent.
Oh wait, they're good at acquiring bonus. I was wrong, they're talented after all.
Whatever happened to hiring the most qualified candidate ???
Whaddja mean "Whatever happened to ..."? It was never there, which is why sensible companies are now making strenuous efforts to introduce it and stop excluding good candidates for irrelevant reasons. No wonder second-raters feel threatened.
Gotta second that, proper diversity pushes are about not excluding qualified candidates for dumb reasons, and expanding the scope of qualifications to include value that that diversity brings to the table. Which can sometimes be quite large.
I was on a project that hired a couple of new junior developers. Turns out that they where multilingual too and could help port the software for internationalization. In two years international sales were 20x domestic sales. (as it turns out there are quite a few people in asia.)
Sadly, that was not a plan by management, it was an accident. Management were neither particularly diverse nor competent, but were great at hiring people just like them.
People talk about "lack of diversity" and there seems to be an implicit assumption that diversity is always a good thing and that something must be done to address the lack of it.
It's actually incredibly hard to take a very diverse group of people and then expect them all to have compatible values, work nicely with each other sing from the same hymn sheet. In the real world, some people are borderline incompatible with each other.
Surprising as it may seem, every project doesn't need (reads from checklist): a man, a woman, a european, an african, an oriental, an indian, an arab, a native american, a rich man, a beggar, a scientist, an artist, an imam, a protestant, a catholic, a nazi, a rabbi, a university professor, a high-school dropout, a war veteran, a hooker, a convicted killer and a bunch of people with rainbow coloured hair. Perhaps if you need a focus group for a product that needs to have appeal to a wide range of people then it's worthwhile, but that is an exception. Most of the time all you actually need is a bunch of people with the right skills who can work well together.
People from different backgrounds see things in different ways, which might help solve a difficult problem, but how often does that actually happen? More likely you find that arguing over lots of contradictory ideas and opinions at every meeting is just frustrating and inefficient. Whereas people with a similar kind of mindset (whatever their background) can often work very well with each other and achieve a lot.
Perhaps we should relabel "diversity" as "lack of homogeneity" and then just hire whoever is best for the job?
of one or many of the criteria you mention which means you will always have a class that is not represented. For example, a black Jewish lesbian person with physical disabilities.
In my opinion what they call diversity is, for all practical purposes, unattainable. With every new binary criteria you add, the number of specimens required to be diverse will have to double. Some category will always be discriminated.
"there seems to be an implicit assumption that diversity is always a good thing"
The aim of diversity policies isn't just to get a different view - it's to reverse the historic "white and male only" recruitment practices which were the norm and, in most industries, still are. In the UK and US a CV with a white-sounding name on it will get more call backs than the same CV with a non-white-sounding name in most industries, including ours.
That the people who can't function well in a diverse team tend to cause problems in homogeneous ones too?
Getting along with people is a skill too, and in my experience, the ones beating the drum about "diversity" being a problem tend to also trend anti-social, difficult, and often outright toxic. Technical skill isn't the only qualification that matters. While some people missframe this as 'better" or "more qualified" employees being pushed out, in fact, it's more likely to push out inflexible and intolerant pains in the ass.
People who still can't get work done aren't going to last long, that's a self correcting problem. People that can, and can work in a team with other people without having a meltdown if they look, talk, or dress differently, will build a stronger and more successful project 9 times out of 10.
Sorry if you seem to be on the wrong side of that job requirement, but that's on you, not on the rest of us.
"Surprising as it may seem, every project doesn't need (reads from checklist)"
I fully agree.
"It's actually incredibly hard to take a very diverse group of people and then expect them all to have compatible values"
A compatible person can come from Europe, North/South America, China, Japan, India, wherever. Dickwad is a dickwad, and nice person is a nice person, no matter where she/he came from. That's my experience so far.
I have never met an HR person who was even marginally competent. At $EMPLOYER they have rebranded their department as "People Services" and the woefully bad thickies who work there as "Partners", and they are still as useless as ever.
They got laid off. They re-hired someone else 6 months later after losing 5% of their core staff, and having open vacancies that were unfilled after two months.
The new HR hired idiots and was sacked after a year and a half. The head of the organization then stepped down, along with several of their direct reports, the equiv of the outfits c-level. The organization used an outside firm to find replacements. They were expensive. 2/3 didn't work out.
The new Head Honcho is now investing in HR as a long term asset.
The people that are running the companies buying and selling this stuff probably aren't even millennials.
On the basis of being a burned out Gen-Xer, I'm guessing we did this at the behest of the same Boomers that used to submit had written applications back in the 80's when they did to much Bolivian marching powder and thought they were cool.
A couple of weeks ago, the FT had an article about the benefits of using AI systems for justice (Jemima Kelly ‘AI-driven justice may be better than none at all’ 29 September 2022, p24 - warning paywall protected). It reckoned that despite the problems, AI justice would be beneficial. I sent a letter in pointing out the many issues (as stated in the Register article and in many of the comments above). The FT published a letter a week later saying that AI justice was actually good, from, ..., a professor of law and AI.
And the people selling this (FT, is it a real suprise?) don't want it to actually be just. It just needs to be plausibly opaque and generate the desired results most of the time. Other "benefit" is it's cheap, and offers a whole new group of people to dip into the coffers in exchange for a new way to put their finger on the scales of justice.
If you want a system that's just, this isn't ever going to be it.
I think I may go throw up a little now. Something smells off.
More management phrenology.
It seems most hr management fads are just ways for high priced consultants to make money, with little or no empiric proof of their value.
This is yet another.
AI is so far away from being useful for this sort of stuff, it's like using a pogo stick to jump to the moon.
Only genuinely conscious sapient AI could do this kind of work. And no one has any clue if that is remotely possible.
If a company interviewed me and relied on fad techniques like AI evaluation or personality phrenology, I wouldn't wait for their decision. I'd withdraw.
If they are that fooled by bullshit at that stage then the culture won't be built with evidence based thinking.
Make sure the department head or nearest equivalent that will be the person's supervisor writes the advert! That same person does a quick sort of the applications received and calls in the best candidates for an interview. The number depends on the time the person has to conduct interviews and the importance of the job. Pick one or two and bring them in on probation. If both work out, hopefully there's enough work for both. If one doesn't, the other should be fine if the job and qualifications are well defined. Most people apply for jobs they believe they can do.
When I had a position to fill, I needed that position filled right futsy now and better is the enemy of good enough. Where I think that all HR fails is they are looking for the best person to fill the role and have screwed up priorities like "diversity" that might get the company served with summons' to appear. As long as the position isn't filled, the job isn't getting done and the company is possibly losing money. Even somebody rather mediocre might be little different than somebody very well suited except the mediocre person may be less likely to jump ship with little notice feeling they lucked into that job.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
