In this issue

• How to protect your data in ChatGPT

• The hidden cost of using AI to slash headcount

• ChatGPT disables Browse with Bing

• Read readers’ minds with this prompt

• Why we’ve created The AI Writer

SECURITY

What’s really happening to all your ChatGPT conversations?

Switch off chat history on all devices, anonymise prompts and protect your password to make sure the data you give AI doesn’t end up in the wrong hands.

There’s one, huge difference between ChatGPT and most of the other tools you use at work. Using it may feel like interacting with any other app (once you get used to the novelty). Yet that feeling can be deceptive.

Contrast it with an Excel spreadsheet, say. You type in numbers and it ‘replies’ with other numbers. But it’s doing that solely with formulae that you or a colleague have given it. And what you type in never leaves your laptop or a secure folder in OneDrive.

ChatGPT is trained not by you but on the 300 billion words that it’s scraped from the internet. (This is the large language model you may have heard of.)

This is one half of what enables it to seem so human, as it knows which words and sentences normally follow others.

The other half? All the information you and its millions of other users type into it.

Company info

ChatGPT still runs on secure servers in the cloud. (The same ones that OneDrive does, in fact.) The key difference is that it learns over time from what you type into it. Tell it that you don’t like an answer and it will try harder next time.

But it’s also banking all the other information you give it. According to British government cybersecurity experts, this doesn’t mean that it’s automatically adding your data to its body of knowledge.

Yet it does mean that the company that runs the chatbot (OpenAI in the case of ChatGPT) can view it and use it to fine tune its models.

This isn’t a secret. OpenAI even says, ‘Conversations may be reviewed by our AI trainers to improve our systems’ when you first log in.

But, in theory at least, the information you give ChatGPT could end up in the answers it gives someone else.

This may already have happened with other user data. In December last year, Amazon warned their employees not to include proprietary information in chatbot prompts after one of its senior lawyers reportedly found ChatGPT responses resembling internal company data.

JP Morgan, Walmart and Samsung have also all restricted its use over similar concerns. Back in March, The Economist Korea reported that Samsung banned it outright after discovering employees had been prompting it with program code and minutes from internal meetings.

Protect your data

But there are ways to protect your data. You can now tell ChatGPT to delete your data after 30 days and not to use it for training the language model.

To do this:

1. Open the side menu (the one with your previous chats in).

2. Scroll down to the bottom.

3. Click the three dots next to your email address.

4. Click ‘Settings’, then ‘Data controls’.

5. Click the toggle switch next to ‘Chat history & training’.

If you’re using the mobile app, you can access settings from the three dots in the top right of the main window.

Important: You need to do this for all devices and browsers on which you use ChatGPT, as the setting doesn’t synchronise across them.

Anonymise prompts

I’d still advise anonymising prompts though, to be on the safe side. OpenAI’s data control policy gives its staff permission to monitor conversations for ‘abuse and misuse’ before the 30-day cut-off even if you opt out of saving your chat history and using your data to train the chatbot.

Third-party apps that are powered by the technology behind ChatGPT are covered by an agreement that means your data will automatically be deleted after 30 days and won’t be used to train OpenAI’s language models. But again, staff can still monitor conversations before then.

Protect your password

Finally, be extremely careful with protecting your password. So far, more than 100,000 ChatGPT logins have been stolen and offered for sale on the dark web. Hackers used information-gathering malware that searches the PCs for sensitive data.

The best way round would be to use two-factor authentication, but OpenAI has recently removed access to the feature, possibly because of technical difficulties. Until it restores it, make sure you use a unique password, change it frequently and keep your anti-virus software up-to-date.

OPINION

Workplace AI will still need humans for a while yet

Think twice before rushing to replace people with bots. The real benefits for organisations lie in getting humans and AI to help each other.

Last month, we reported on how some organisations were using AI to cut headcount. At least 4000 job cuts in May were blamed on AI, a charity fired its entire helpdesk and replaced them with a chatbot in the same month, and IBM is planning to pause recruitment for many non-customer-facing roles.

This is obviously bad news for the staff affected. But rushing to replace people with bots could also be bad for business and the overall success of any organisation.

Why? It’s too simplistic and too early.

Real wins

The real wins from AI will come not so much from the wholesale replacement of humans but by enhancing the work that humans do. The research that we previously reported on how a generative AI conversational assistant improved customer service was a good example of this.

Bots were able to suggest solutions to agents in real time while their human counterparts were deep in conversation with frustrated customers. The agents could then choose to accept or reject the solutions, many of which may not have occurred to them.

The productivity and customer-satisfaction gains organisations could make if every customer service worker had their own virtual personal assistant should be obvious. But the study showed that it was the least-experienced (and presumably cheapest) employees who benefited most.

AI customer-service assistants slashed the time that new hires took to get up to speed in their roles.

But rushing to replace humans with bots also risks squandering the chance to reap its biggest rewards. Up to this point, most organisations have depended on people, with all their imperfections and human foibles.

Scaling bad practice

It follows that the work those people will have done will have been limited as well. If you replace those people with bots, then you scale up their imperfect work too.

AI may do the work more quickly, but the gains are always going be limited if they’re doing more of the work that was always suboptimal.

A far better way to make the most of the AI revolution is to use it to help humans improve the work they do. This will release them from limitations that have always held them back, freeing up the time and brain space to find a better way of doing things.

We can then use AI to scale that.

NEWS IN BRIEF

ChatGPT disables new web browsing feature

Reports that users had been using ChatGPT to bypass paywalls have led OpenAI to switch off Browse with Bing.

OpenAI has temporarily shutdown ChatGPT’s new web-browsing feature just a week after it went live after finding that it was a little too effective.

The ability to access the web was one of the most significant additions to ChatGPT since it launched last November.

But a statement from the company posted earlier this week said, ‘We have learned that the ChatGPT Browse beta can occasionally display content in ways we don’t want. For example, if a user specifically asks for a URL’s full text, it might inadvertently fulfill this request.’

Translation: Users have been using ChatGPT to bypass paywalls. 

Oops.

Web browsing was introduced recently as a beta function, meaning it was still in development and made available so OpenAI could get user feedback.

Only subscribers to the paid-for version of the AI bot, ChatGPT Plus, could access it. If you have a Plus subscription, you can still access the web through plugins such as WebPilot and LinkReader.

HOW TO

Read your reader’s mind with ChatGPT

With this prompt, AI can help you regain one of the most effective, human writing qualities – empathy.

Why you need this

When documents and emails fail, the blame often lies with a lack of empathy, at least in part.

It’s not that the people who write them can’t empathise. But we get so tied up in our busy, often stressful lives that projecting ourselves into the position of the person we’re writing for can be incredibly difficult.

It doesn’t help that writing itself is a solitary process. Alone with a laptop and a full to-do list, it’s only natural that we focus mainly on ourselves and our own priorities. We’d be much less likely to do this if we were speaking to the other person.

This is often why our reports, proposals and emails can land flat or even irritate the people we send them to. (It’s also one of the many differences between speaking and writing that hold us back now we’ve become so dependent on the latter, as I explained in this article.)

It’s a double-edged sword. What we’re writing makes perfect sense to us precisely because it’s focused on us. So it can come as a real shock when we don’t get the positive reaction we were hoping for.

Why this works 

Surprisingly, AI can help us regain our empathy, even though it’s the most human of qualities.

To be clear, AI can’t ‘think’, no matter how much it may feel like it can. But this also means that it can't focus on its own needs, for the simple reason that it doesn’t have any. It’s totally objective in the way that we aren’t.

To get the best from it, though, it needs a nudge to remember that the people we write for are just that – people. (We all need reminding of that from time to time.)

The prompt below is simple but powerful. It’s a great way to increase your chance of a positive outcome by shifting your focus from your needs and priorities to those of your reader.

What to do: 

Copy and paste this prompt into ChatGPT or GPT-4.

I am writing a report on [topic]. The report is for a [role] in a [organisation/type of organisation] in [location]. What aspects will most appeal to their individual human interest? Please present the results as a bullet list.

Here’s an example, using GPT-4:

Note: To tailor this further, try adding ‘They are also particularly interested in [X, Y and Z]’ to the prompt.

Why The AI Writer?

When I founded Emphasis, 25 years ago, it was just me, a cat and a kettle. It’s since grown to become the most trusted provider of business writing training in the world and helped more than 80,000 people, from 32 countries.

We’ve worked with tech giants, top 10 law firms and major financial institutions, as well as at the highest levels of government. (We’ve even sent trainers to work with clients in the Himalayas.)

I would not claim that we’re experts in everything, as we’re certainly not. But, after working with the authors of around 100,000 documents, it’s safe to say that business writing is something we do know a thing or two about. And that includes witnessing all the various ways in which organisations get it wrong.

Now, we’re at the forefront of using AI to help them get it right.

Of course, I’ve seen nothing like the AI revolution that’s hitting us. Nobody has. But I do feel a responsibility to use our experience to help people navigate this brave new world of communication bots.

So here’s the deal. You get on with your job, while we immerse ourselves in the latest developments in workplace AI.

We’ll worry about keeping up so you don’t have to. Then we’ll tell you exactly what you need to know to stay ahead of the curve.