It was a warm, breezy Saturday morning in Melbourne, two years ago. Michael Fowler had just shut down his laptop after spending a few hours checking up on his business. He was the owner of a T-shirt company called Solid Gold Bomb, which sold a wide variety of garments online via, among other outlets, Amazon. Fowler was looking forward to getting back to his weekend, but then his phone buzzed. A notification from the company’s Facebook page. It buzzed again – and again. When he went to read the messages he saw a stream of abuse directed at his firm. The memory is still clear in his mind. “I knew instantly what had happened when I read what people were saying,” he says. “I couldn’t reply fast enough.”
Fowler had set up an algorithm to upload thousands upon thousands of T-shirt designs to his online stores. The designs were based on the infamous “keep calm and carry on” catchphrase, a slogan which was originally dreamt up as a way of preserving morale in the event of a Nazi invasion of Britain. Fowler had decided to “parody” it by getting a computer programme to come up with random variations such as “keep calm and dance on” or “keep calm and play football”.
Fowler’s employees were all out of jobs and a once thriving firm was gone. All because of a horrible T-shirt that no-one wore, no-one bought, and which never materially existed.
But the huge list of word choices that he fed in included less savoury options – which he says he had no knowledge of. In particular, a T-shirt emblazoned with the imperative “keep calm and rape a lot” had been published. No-one had bought it. In fact, it sat on the web for more than a year before anyone even noticed it. But eventually it was discovered – and the internet went crazy. Twitter was ablaze with condemnation. “Solid Gold Bomb is crouching behind its algorithmically generated excuse,” said Gizmodo. Others pointed out the stupidity of using an unexamined word list to automatically generate slogans for a commercial product.
Fowler admits he made a “big mistake” and within a couple of months, Solid Gold Bomb folded. His employees were all out of jobs and a once thriving firm was gone. All because of a horrible T-shirt that no-one wore, no-one bought, and which never materially existed.
But that’s the trouble with algorithms. All sorts of unexpected results can occur. Sometimes these are costly, but in other cases they have benefited businesses to the tune of millions of pounds. What’s the real impact of the machinations of machines? And what else do they do?
To really understand the sorry mess that ended things for Solid Gold Bomb, one has to consider the instant success that Fowler’s scripts and computer tricks had once brought him. Years before he ever experimented with the “keep calm and carry on” meme, he had devised an automated T-shirt design process which published over 22 million different versions of sports-related designs to a web store. These included icons and, crucially, people’s names. Finding the shirt with your name or your friend’s name on it made you much more likely to buy it, discovered Fowler. “It was about a 100-to-one ratio. For example, a picture of a car would sell once whereas a picture of a car with a name below it would sell a 100 times,” he says.
I’ve realised that you have to have an element of scrutiny – Michael Fowler, online seller
When Fowler first launched this technique in 2011 he received 800 orders over the first weekend. He was blown away and the effect on his business was profound. Soon he was processing thousands of orders a day. The problem was that a huge proportion of the designs weren’t vetted for suitability by human eyes before they went live on the web – an oversight that would lead to the problems two years later. “I’ve realised that you have to have an element of scrutiny,” he admits.
Fowler has now returned to selling T-shirts – after a spell as a traffic warden and as a ranger catching stray dogs – but today he is more careful. His current company, Big Texas, also uses an algorithmic process to create designs for aprons, but he uses published lists of the 1,000 most common names, not random assortments of words.
Errant algorithms can also cause human headaches when it comes to prices. The costs of products that appear on retail websites are constantly fluctuating thanks to software that sets them competitively. The frequency at which these changes happen is so great that dedicated websites have been set up to “watch” the pricing on websites like Amazon. Daniel Green has been running one of these sites for years. He explains that prices don’t just change daily – but sometimes several times in one day.
Prices don’t just change daily – but sometimes several times in one day
“They will drop the price of a product every few days or every few hours until a product is purchased by someone and then the price goes back up,” he says. “We know that they keep prices low on a lot of their most popular products to give the impression that they have great deals and then for less active product categories or less popular products they may have a bit more of a profit margin there.”
Sometimes this can produce amusing and unexpected results, however, in what Green calls a “race to the bottom”. Two retailers selling the same thing on Amazon’s marketplace will re-price their product against their competitor, but the re-pricing can occasionally continue unabated until absurdly low or high price points are reached. “It just goes back and forth,” says Green.
In 2011 a blogger discovered that a biology textbook about flies was priced at the bewilderingly high price of over $23 million. Why? Two sellers, profnath and bordeebook, had set up algorithms which would watch each other’s prices and then reset their own. “The prices would remain close for several hours, until bordeebook ‘noticed’ profnath’s change and elevated their price,” explains Michael Eisen on his website. The pattern continued perfectly for the next week.”
Algorithms watch other’s prices and reset their own
Sometimes the most damaging prices are the lowest – as low as you can get, for example. This was what blighted the Christmas sales of a string of Amazon marketplace sellers in December last year, when automated re-pricing software RepricerExpress erroneously changed the price of thousands of items to as little as one penny or one cent. The glitch, which happened two weeks before Christmas, hit sellers running small retail businesses extremely hard.
Part of the trouble was that once orders had been placed, they were almost instantly processed by Amazon’s automated warehouse systems under the “Fulfilled by Amazon” (FBA) programme. Sellers asked the FBA team to stop sending these orders out but were told it wasn’t that simple. Amazon says that nothing about the operation of this system has changed since the December incident.
It almost felt like somebody broke into your house - Richard Burri, on an algorithmic glitch
“It almost felt like somebody broke into your house or your personal life and started to take things away from you,” says Richard Burri, whose office stationery store was affected by the error. He and his wife estimate that the various computer algorithms working together would have cost the business between $100,000 and $150,000. Fortunately, the majority of the firm’s human customers who had bought one penny items agreed to return them when contacted.
Others found that buyers weren’t always so obliging. Shamir Patel sold pharmaceutical products via Amazon. He also asked customers to return one penny products, but he says about half of them refused to do so. The cost to his business, he calculates, was around £60,000. “You were a bit powerless to do anything about it,” he recalls. “You were literally just watching your money flush down the drain.”
It’s made me not really want to go on with the business – Judith Blackford, online seller
Yet another seller, Judith Blackford, who marketed dolls and babies’ gifts, says she suffered sleepless nights and adds that she has been forced to close her business completely. “I’m selling off any stock that’s left now. Pretty much most of it is going out half price just to get rid of the stock,” she says. “It’s made me not really want to go on with the business.”
RepricerExpress has apologised for what happened, though the firm has never explained exactly what went wrong. CEO Brendan Doherty declined to provide comment for this article.
All of these stories demonstrate how acutely online retail is defined by the machinations of decision-making software hiding behind the scenes. For Michael Fowler, he learnt the hard way that algorithmically derived success can quickly turn into spectacular failure. “You can scale up and have hundreds of orders overnight. Hopefully you’ll have a process to know how to do that but I think that’s the biggest problem – you have to be ready for it and know how to handle it,” he says, reflecting on his experiences. “I don’t think it’s a long-term, sustainable solution unless it is really properly executed and done conservatively – with a lot of auditing in place.
“If I wanted to I could scale my current business up tomorrow but I think I just have a more conservative approach now. I’m just trying to survive.”
Follow us on Facebook, Twitter, Google+ and LinkedIn