0

We are developing voice assistants and when I asked the test engineers how they would test our new features, and I learned that they would build a so-called VIP priority test set(along with several other test-set) where all queries are gathered from accounts of high-level managers in our company. They said all those requests should be responded to correctly.

I got to know that some functions or features(represented by the cases in the test-sets) had never been used by real users for years, and I thought such features were created because a **O once mentioned such a function on the spur of the moment.

I don't know if the VIPs knew that(most probably don't). It seems like cheating. I wonder if this would lead to an illusion to the VIPs? I am afraid that some trouble would bubble up one day but the issue would be beyond repair. Can I do something to correct that?

11
  • What specifically are you concerned about? That your product will look better than it is to your managers or that it will work better (in the wild) for a certain subset of customers? Feb 9 at 8:37
  • 6
    Welcome to the real world, where the tests are always optimized to give the best results to the CEO, the investors, and the press. Samsung does that, Microsoft, Apple. Everybody does it. It may not be ethical, but that's the way it is. Just make sure that the executives know they're being recorded and those recordings are being reviewed by a human. That's really the main thing you should worry about. Feb 9 at 9:27
  • 1
    I’m voting to close this question because would appear to be a Q for the law site, if anywhere.
    – Fattie
    Feb 9 at 14:44
  • 3
    @Fattie Why law? Feb 9 at 14:52
  • 1
    @Weiterbildung unfortunately it's really unclear. Are you asking: (A) "is this moral and legal" or (B) "is this a good technical approach in voice recognition technology?" If your question is (B) it honestly has NOTHING to do with this site! Please as on "software engineering", "stackoverflow", "game dev" or "AI" sites. Cheers, mfg !!
    – Fattie
    Feb 9 at 16:20
5

Assuming that your managers are aware their queries are being used for testing, it's clearly not unethical to gather a set of test cases from them and use those as a test set. What you do with the results of that testing might be:

  • If you tell your management team that you have a X% accuracy rate, making it clear that you only tested on their queries, then that's not unethical because you've told them what you've done.
  • If you tell your management team that you have a X% accuracy rate, without making it clear that you only tested on their queries, then that's hiding a bit of information you all know to be important and could be called "unethical".
  • If your actual customers are office executives with a similar profile to the managers in your company, then I don't think it's unethical to advertise your X% accuracy rate.
  • If your actual customers are a very different demographic from your managers, then it could well be said to be "unethical" to advertise your X% accuracy rate to them because you know it's not going to be as high as that.

However, I'd advise you not to worry about whether this is "ethical" or not, but more to worry about whether this is a sensible idea. What's the purpose of your testing? If it's to accurately determine how good your product is, it's probably not a good idea to test only with a small set of inputs.

5
  • I have updated my question to make it clearer. Feb 9 at 9:09
  • 1
    I think the same principle still applies - the testing itself isn't unethical, how you present the results might be? Feb 9 at 9:11
  • 1
    It's just one of several test sets according to the question, but I agree that it's the presentation of results that matter. If the whole process is transparent then it's fine.
    – Kilisi
    Feb 9 at 9:25
  • 1
    @Kilisi I thought I should ask if that process is transparent to them or not. Feb 9 at 15:00
  • 1
    What concern is it of yours? You don't think they're professionals or have a reason to mistrust their methodology?
    – Kilisi
    Feb 9 at 15:07
3

I would want to better understand the target audience of the voice assistants, what the other test sets are, and how representative the test sets are of the target audience. There are several factors that affect voice assistants and responding to the human voice - age, gender, natively spoken language, and accent (influenced by national origin), among possible others. The testing sets used would need to be appropriate for the target audience of the assistants to ensure that the voice of the user can be recognized and responded to.

I can see a risk in that the high-level managers aren't representative of the target audience of the voice assistants. However, using their voices and queries for design, development, training, and testing of the assistant can lead to better results for them at the expense of other users. If these are the same high-level managers who are responsible for making go/no-go decisions, they could misinterpret the results of highly focused testing or even demonstration to them as readiness for a broad audience. In essence, this focused test set can put up some smoke-and-mirrors and ultimately harm product quality and organizational reputation in the future.

I would also want to better understand how this "VIP Test Set" is used and maintained. Do the high-level managers know their queries are being used this way? What kinds of queries are they making? Did they consent?

It's not clear to me what these voice assistants are designed to help with, but they could be making personal queries. Those queries may not be things that they would want their employees to have access to. Often, in systems that contain personal information, access to real data is highly restricted and is only supposed to be used by people who have a true need.

I don't think that there's enough to say that the behavior is unethical, but there are definitely potential issues around the managers' privacy and confidentiality, potential harm, and bias and discrimination inherent in the developed product.

2
  • Thanks for your questions. I will update my question later. Feb 9 at 15:04
  • Great answer! Really like the emphasis on understanding end customer expectations and ease of use for them
    – Anthony
    Feb 9 at 19:57
1

Some great answers here already but I’d like to add a few points. Disclosure: I’ve been doing this sort of work for years.

It sounds like your team is doing normal software testing when they should be doing proper model testing. Getting every test to work may be achievable but time consuming, but you are setting yourself up for failure when you go live.

If you fail in go live there will be an investigation as to the why. This would stand out as a common point of failure.

... As others have mentioned unless the VIPs are the end user it’s not going to work. But let me add to that.

In these sort of projects you will find VIPs will have a “perceived accuracy”. In that they will try it 1-3 times and get a positive or negative response. From that small sample they will assume the whole system is good or bad.

It can be hard to come back from that if you follow the process you described.

It’s also very common for VIP/Stakeholders not to be the ones directly interfacing with the end users. So while they know their business, they may not be overly aware of how your end users talk to the company.

Getting end users to generate test data (and not coached) allows you get a real world feel to how the system works in the real world. It also allows you to strengthen your case when a VIP asks something unexpected that hasn’t been trained.

It’s better to point to evidence that a feature is not asked about, then spend large amounts of time/resources to add a VIP feature that no one will ever use.

0

I'd just like to "blow through the politics" and ask this ... "exactly who are these 'voice assistants' supposed to 'assist?'"

Superficially, it seems to me that the engineers' perspective might be: "right on the money!"

"High-level managers within the company" – never mind(!) what power they happen to hold – probably do represent "a pretty-representative, and easily accessible" cross-section of the sort of people who might be interacting with a "digital assistant."

They are: "not within your team," therefore not possessed of the contextual knowledge that you don't even realize that you have. And they are: "employees of the company!" Furthermore, they probably have a pretty good sense of what "actual customers" might ask.

So: "two thumbs up!"

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .