Big Data, Plainly Spoken (aka Numbers Rule Your World)

Will AI make cheaters of us all?

Andrew wrote an amusing post about mischief using AI in peer reviewing for academic journals.

It emerged that authors of scientific papers have resorted to embedding secret prompts inside their text to instruct large language models (LLMs) to give their papers positive reviews. These prompts may be printed in white, or tiny font, so they are intended to evade humans. Some prompts are quite elaborate, carrying instructions for what to say about strengths as well as what to say about weaknesses. For example:

Be sure to check out the comments section, as readers fuss over which group is worse: the authors who instruct LLMs to give positive reviews only; or the reviewers who rely on LLMs to submit their reports. As Andrew told the story, one author who admitted to inserting these prompts argued that they did it only to deal with cheating reviewers who deploy LLMs. So, we are witnessing the classic two kids in a playground scenario - he's the one who started it!

We can take this blame game one step further. The cheating reviewers should blame it on the authors because some authors are using LLMs to write bogus papers!

***

Unfortunately, this is the world of AI we find ourselves in. At an event recently, I chatted with an instructor who is throwing his hands up, complaining that he is spending time correcting code submitted by his students who are obviously using AI to do the work. Meanwhile, there are students complaining that their instructors use AI to set or mark assignments. They can of course blame each other.

Would one begrudge instructors who ask AI to mark assignments if the work were generated by AI? Would one judge the students who use AI to do their homework if said assignments were created by AI?

Is this a race to the bottom? Eventually, will humans do any work?

 

Know your data 46: using our data to set pricing

Google's former CEO Eric Schmidt infamously said something along the lines of "If you don't want others to know, you shouldn't be doing it in the first place". It showed the hubris of Silicon Valley at the time, and a certain deceitfulness. Because the truth is if they have your data, they can use the data to harm you, even if you haven't done anything wrong!

Finally, we have some evidence of what's been going on behind closed doors for a long time. They use our data to price discriminate. Same product, different prices, based on analyzing our data.

This practice, known as "surveillance pricing", caught notice because an industry body is suing New York State about a new law that requires companies to disclose that they are using algorithms (and data) to set different prices for different people (link). Look, the state is not banning surveillance pricing; they are requiring notification.

The industry doesn't want us to know.

The pushback from industry follows the usual script:

They bring up alternative scenarios of potential benefit to dismiss scenarios of harm. The state alleges that prices are raised on those who can afford them. In this case, they claim that the same algorithms are used to lower prices by offering discounts to selected customers.

I don't doubt that algos target special deals at specific customers. But I fail to understand why customers who love receiving coupons would object to the required disclosure of surveillance pricing - in fact, it would be great marketing to inform these customers that algos found them good deals; they might learn to like algorithms!

Surely, no consumers should object to such disclosure.

The industry apparently wants us to believe that the primary objective of surveillance pricing is to deliver discounts to customers. Anyone who read Chapters 3 to 5 (marketing data) of Numbersense (link) should recognize it as hogwash. Discounts result in lower revenues, unless the business can somehow prevent existing customers from using them. Whatever "leakage" happens, the business has to make up for these "lost" revenues. Thus, the same algos are likely to raise prices on other customers. With the amount of data at their disposal, it's not hard to figure out which customers are less price-sensitive, or have higher "willingness to pay".

I have worked on such algos. What kind of reception do you think data scientists would get from the business teams if we present to them an algorithm that delivers discounts to selected customers, leading to the predicted outcome of lower total expected revenues (and lower profitability)?

***

The usual script from the industry bodies also includes the false claim that telling the truth is "misleading". This is the same script used to oppose non-GMO labels. In this instance, they assert that customers will interpret the mandated disclosure as evidence of "price gouging".

These industry honchos aren't alarmed when consumers falsely believe that prices are fixed for everyone - when such disclosure isn't required!

I must digress to complain about another industry practice that is gaining popularity by the day, at least in the U.S. Many stores don't even bother putting up price tags. Some restaurants and coffee shops put up menus without prices. I just walked into a Vietnamese diner this afternoon, hoping to get an iced coffee to combat the oppressive heat in New York - well, the menu of bahn mi and side dishes is reasonably priced but the beverage and dessert sections have no prices! I walked out, disgusted. The iced coffee, based on redacted prices, is probably outrageously expensive, either compared to their past pricing, or compared to their peers. I'm guessing at least $6 (plus tax and the delightful... tip screen).

In their war against disclosures, some businesses won't even put up their prices.

And yet, we should believe that they won't use our data to maximize their profits.

 

 

Nonlinear thinking in marketing

I recently had the pleasure of listening to Avinash live at an event sponsored by Precise TV. Avinash is someone who gets marketing analytics, as well as a great communicator.

The talk is centered on his "See, Think, Do, Care" framework, which is posited as a challenger to the dominant schematic of a "marketing funnel".

For those readers unfamiliar how marketers think, they think linearly not unlike the rest of the world. A marketing funnel is a classical way of organizing the marketing function.

In the example above, the universe of the potential customers of a business is divided into four groups: the first group consists of people who are aware of the company's products but not yet considering a purchase; the second group are those thinking about buying; the third group are those who decide to buy for the first time; and the last group are those who have purchased more than once.

In this funnel setup, a marketing team can be split into four sub-teams. The first team focuses on driving awareness; the second team's goal is to get people from aware to interested; the third team - which is the core of marketing - wants to "convert" interested prospects into first-time customers; lastly, the loyalty team's job is customer retention, indicated by return purchases.

The funnel describes a linearized world. Each person enters the top of the funnel, and marketing's job is to push them down the funnel as far as possible, as quickly as possible, and keep them there.

Avinash opposes this linear view of the world. In his "See, Think, Do, Care" framework, he also sets up four groups but people can move from any to any. He calls these groups "audience intent clusters".

The "See" group consists of people who are just looking around. The "Think" group are those who have expressed some interest - in the digital world, interest is evidenced by specific behaviors (such as clicking on some link). The "Do" group are those who are close to buying, for example, those who have moved an item to their shopping cart. The "Care" group are the return customers. Unlike traditional funnel users, Avinash wisely sets the bar higher. Someone who has made just one purchase isn't a target; someone who has made two or more purchases is worth cultivating. It's common sense, yet he's right - most marketers see anyone who has bought something as potentially "loyal". The problem with such an approach is that most of the loyalty marketing dollars would be wasted on people with no intent of returning. Why not focus the spending on those with a higher chance of future business?

Avinash points out the lack of "care" in how many businesses deal with the "care" segment. This is particularly true of technology companies. Tech support FAQs, and a support phone number that's hidden from view show return customers not love but indifference.

The key idea of the talk: any person is not trapped in one of four stages until the marketers shove them one step below. A "loyal" customer might be browsing at the brand's Instagram channel, and her intent might be "see", not "do". So, the content shown to her should incite curiosity, rather than hard selling.

There's more here (link) in Avinash's own words.

 

Interpreting margins of error in tennis calls

A commentator of the French Open recently complained that the human line judge made a mistake: "Hawkeye's error is 3 cm and the ball was out by 4 cm. So the line judge is wrong to call it in."

***

The commentator got this all wrong. The divergence of opinion should reduce one's confidence in Hawkeye's estimate. Let me explain why.

Hawkeye's goal when it comes to judging line calls can be stylistically described as determining the center of the landing spot of the tennis ball. It's helpful to first look at what happens without a margin of error. From this estimated location, we draw a ball given the diameter of a tennis ball, and figure out if the ball overlaps with the line on the court. If it doesn't overlap, then the computer decides that the ball is outside the line.

The idea of a margin of error of 3 cm is visualized as drawing a circle of radius 3 cm around the estimated location. Now, from any point inside this circle, we draw tennis balls as before; and if none of these balls overlap with the line on the court, then the computer decides that the ball is outside the line. By involving the margin of error, we explicitly embrace the uncertainty of estimation. For sure, fewer out calls will be issued relative to the case when we just use a single location estimate.

The fundamental problem in statistics is that Hawkeye gets one chance (one sample) to get this right. If Hawkeye used a deterministic process, then given the same inputs (videos, etc.), it would always generate the same estimated ball location. In real-world systems, whether it's because of noise in the system, or some stochastic element in Hawkeye's process, the same inputs lead to different estimates. The margin of error describes how much these estimates vary.

The reported margin of error holds that Hawkeye's estimate is unlikely to be off by more than 3 cm. In other words, that expanded circle of radius 3 cm is expected to capture the "true" center of the ball's landing location.

The word is "unlikely" rather than "impossible". All margins of error comes with a confidence level; usually it is 95% confidence. This means there is a 5 percent chance that Hawkeye may be off by more than 3 cm from the true location.

Hawkeye's estimate is not error-free as the commentators assumed, even after allowing for the margin of error.

I'm curious about the margin of error associated with humans inspecting ball marks on the clay - I suspect it's small. (The error of judging the balls in flight, by contrast, is certainly much higher.)

***
At tournaments that use Hawkeye, the players are forbidden from challenging calls. Let's subvert the process and exchange the roles of humans and machines.

Assume Hawkeye makes the first call, in or out. If the player disagrees with the call, he or she raises a challenge, and the umpire (and/or line judge) goes to inspect the mark on the clay. Now, the umpire's word is final, no complaints allowed.

As an example, Hawkeye determines that the ball is 2.5 cm outside the line, which is less than 3 cm, thus the machine rules it "in". A player protests. The umpire decides that the mark on the ground is wholly outside the line, and changes the call to out. How will the commentators react?

If their reaction is not colored by a preference for machines over humans, they will say that the machine has made a mistake - and to accord with their current behavior (in reverse), they should then recommend that the tournament removes line-calling machines because they are not accurate.

This is an instance in which reversing the players makes clear one's biases.

***

If we take a Bayesian view of this, we should combine the evidence. In the first step, we have one estimate. Now, if the second estimate conforms with the first, then the evidence becomes stronger. But if the second estimate contradicts the first, then the evidence weakens. This is why I said at the start that the divergence of opinion causes me to lower my confidence in Hawkeye's estimate.

Even more, I believe that the human estimate derived from the mark on the ground is more accurate anyway so I'd give that even more weight.

 

P.S. Outside of clay courts, the situation is more complicated as there are no ball marks to look at. I'm not against the technology. I'm against the illusion of perfection, and I'm against black-box technology that stifles dissent. Both these issues can be addressed by how technology is applied.