Why Good Writing is Central to Conversational AI
What role does good writing play in the day-to-day business of bot building? And what is “good writing” anyway? In this article, we explore some of the ways good writing plays a crucial role in your conversational AI strategy.
By Rob Lubow
December 6, 2018
Saying good writing is central to conversational AI is a little like saying that good paint is central to painting, or that cute, furry animals are central to the petting zoo experience. After all, the building blocks of conversations are words. Therefore, the words should be high-quality, right? It seems obvious, but with so many new AI technologies and compelling use cases to implement, it’s all too easy to lose sight of the importance of good writing in the bot design process.
Granted, this need for human writers may change someday. Advances in natural language generation (NLG) – where the bot generates responses on its own – are a few years on the horizon. That’s fantastic news for the future, but for your bot to win hearts and minds today, a good writer is an essential piece of the puzzle. But what’s good bot writing, exactly?
Getting the words right
With few exceptions, all good writing tends to share three qualities: 1) clarity, 2) concision, and 3) appropriate tone. If your bot has these three fundamental strengths, it’ll make an excellent first impression and save you months of optimization. By contrast, failing to possess any of these three qualities could damage the credibility of your bot, and by extension, your brand.
The key to clear writing is getting the grammar and spelling right. Not because you want an A on your report card, but because if users don’t know what your bot is saying, what’s the point of having a bot in the first place? Here’s an example of poor grammar leading to unclear writing:
Botcopy Chatbot: The folder was on the bus, but now it's gone.
Which is gone, the folder or the bus? Grammatically speaking, the bus is gone. And yet you have an uneasy feeling that maybe the bus is still here, and the folder’s gone. The sentence’s ambiguous pronoun – it’s – has you confused, which isn’t cool, especially if it’s your folder.
To avoid unclear antecedents and a host of other grammar gaffes, use a grammar check app or at least get the help of a human grammarian. Better yet, do both. Grammar apps in the hands of non-writers can be of limited value, as we’ll see below.
Next up, concision
now that we’ve got the subject of clarity squared away, let’s move on to the critical topic of being quick and concise.
Cutting unnecessary words is a rule of thumb in all writing, but it’s vital when writing for bots. See below and consider which bubble has the best odds of keeping your attention. Can you guess which was written by a pro?
Botcopy chatbot: Okay, now that we've got the subject of clarity squared away, let's move on to the very important topic of being quick and concise.
Botcopy chatbot: Okay, on to the topic of being concise.
Botcopy chatbot: Now, a word on concision.
Botcopy chatbot: Next up, concision.
Okay, which was written by a pro? All of them. That’s because clear and concise writing isn’t written. It’s edited. The process typically goes like this: 1) write wordy copy on a bot building platform and tell your team that the bot is good to go, 2) actually see the text on your smartphone for the first time when your bot goes live and realize that you wrote way too much copy, 3) frantically run to your computer to shorten the copy, and 4) repeat steps 1-3. Forever.
The elegant brevity of the copy in the last bubble above isn’t something a grammar app can achieve on its own. (Yet.) A human needs to be involved, preferably one with the skill, judgment, and vigilance (OCD) to ensure your bot speaks in smooth, snappy dialogue most of the time, without compromising on the brand’s objectives.
Having a writer like that on your AI development team is essential. Data scientists are generally dispassionate about writing style because they’re confident they’ll arrive at the winning sentence based on iterations over time. Their confidence is warranted, but it’s no excuse to publish dreck at the get-go. Your bot’s first users may very well be the most important – they’re among the rare group who managed to discover your bot and are willing to chat. Don’t blow this golden opportunity by treating these people as guinea pigs.
If you have clarity and concision, you’re ahead of the game. But, that’s not enough. You don’t just want users to use your bot. You want them to like your bot, so that they’ll stick around longer, spread the word, and keep coming back. For that, you need a likable tone of voice.
The tone is the x-factor that makes your bot feel familiar, friendly, and appealing. While grammar tools can somewhat assist with clarity and concision, you’re on your own with the tone of voice. There’s no algorithm (yet) that can generate the exact right tone of voice that feels authentic to your brand and resonates with the proper demographic. Writing like that requires great writers.
In the example below, let’s pretend the user typed “Set up a meeting.” It’s a standard query answerable in many ways and the answer usually has a button under it for the user to tap and trigger a scheduling feature. Below are some of the various ways your bot can answer. In each example, how does the bot’s tone make you feel about the brand?
Note the vast difference between the first and last options. There’s nothing inherently wrong with any of the messages, except that the first one is a bit tone deaf and overwritten – for instance, you don’t need to say “tap the button below” because it’s evident to most users. Also, real people use contractions and rarely scream. Thus, that is a great idea! would be more natural as that’s a great idea. Digging further, do we need to tell the user their idea is great? There’s nothing wrong with affirmations, but if you answer too many user requests with great idea, cool, wonderful, you risk saccharin or desperate tone.
Get this party started is grammatically correct and concise. It might be perfect for a wedding planner bot, but terrible for a banking bot, unless it’s for an uncommonly fun-loving and irreverent bank, which could easily be the case. No prob is informal to the extreme, and again, could be quite good or wholly bad depending on the situation.
All of the hair-splittings above are matters of tone. These are precisely the kind of nitpicky observations good writers make and have always made, long before bots hit the scene. That’s why copywriters or creative directors – not the programmers or UX designers – are most qualified to serve as Queen’s Guard to your bot’s tone of voice.
That said, anyone can spot a tone-deaf line. Often, it’s a non-writer who points out that a swatch of dialogue “sounds bad.” When that happens, it’s the writer’s job to futz with the line until everyone agrees it “sounds good.” Only then is it ready to be tested in front of a live audience.
The unique challenge of writing for bots
Today’s best and most experienced writers began their careers writing brochures, print ads, billboards, TV and radio spots, and later, web copy, email marketing copy, social media posts, and longer content articles. These artisans bring clarity, concision, and tone of voice to the table. But that’s not enough. For their writing chops to be valuable, they must complement the complex process of bot development.
Non-linear journeys & training phrases
Instead of writing linearly, bot writers need to think tangentially, discursively, and in fractals. To explore what that gibberish means, let’s take that last sentence and put it to the test in the mouth of a bot and see what happens.
Botcopy chatbot: Instead of writing linearly, bot writers need to think tangentially, discursively, and in fractals.
User: what the heck does that even mean.
Exactly. Whatever bot writer had the gall to write that bot’s sentence about fractals and stuff better be prepared for the kind of user retort above, along with wtf are you talking about? / huh? / in english pls, and so on.
To train this bot to understand these kinds of user retorts, the bot writer will have had to pre-program the above utterances, and many more, as training phrases. With many platforms, machine learning will kick in and help the writer flesh out the list of possible follow-up retorts, but the lion’s share of the endeavor often requires significant amounts of empathy, and linguistic virtuosity on the part of the writer, such that the bot is prepared to respond to wtf does that even mean with something like:
Botcopy chatbot: Bot writers have to plan ahead for the conversation to take twists & turns. They have to be ready to unpack statements, and unpack the unpacked statements, and so on.
Well played. But the writer’s job isn’t over yet. Not even close. Because now the user wants to know something else:
User: what r fractals?
Oops. The bot writer in this hypothetical example will likely think long and hard before ever again scripting compound sentences containing superfluous twenty-dollar words like fractals. But below we see that the brave writer planned even for this eventuality with this follow-up:
Botcopy chatbot: Fractals are like branches that branch into small branches, and each of those smaller branches branch into a branch of smaller branches, and so on.
This answer would also be triggered if the user asked in the form of explain fractals / what does that third word mean / fraktols?! and so on.
At the time of launch, the training phrase bank for this single piece of packaged data above might be 50 utterances long. A professional writer is an ideal person for training bots with a variety of user queries because writers 1) can type fast, 2) have an ear for plausible wordings and idioms, and 3) are accustomed to rapidly generating synonyms and permutations during the editing process.
Next time, the writer in this example will also want to avoid open-ended statements, and instead get users back on topic, like this:
Botcopy chatbot: It's like a branch that branches to a smaller bunch of branches, and so on. Now, what can I answer about product X?
It’s not easy to manage tangents, but when writers have the creative freedom to design great follow-up arrays and training phrases, bots can suddenly take on new, unexpected dimensions. They stand out, and seem truly conversational, in that they can answer any user request that is reasonable within the context of the bot’s stated purpose.
Again, many tools save writers time in training by creating common training phrases, e.g. yes, yep, sure thing, and so on. Over time, machine learning technology fleshes out the permutations and scores them based on statistical probability. Also, bot writers may need to set up and use entities to gather information, such as location, order name, product, etc. The best bot writers have a handle on setting up entities and including them in bot training phrases.
Putting words in the user’s mouth
In addition to response and training phrase copy, there’s a third component that requires expert writing. Menu buttons and quick replies are a great way to suggest options to users and keep them on a sensible path toward their goal. When these buttons are well-written, they do a lot more than serve as straightforward options – they become attitude adjusters.
Imagine your bot has just introduced a new product or service to a user. In the hands of an average writer, the copy in the blue button below might read “Learn more.” Nothing’s wrong with that – I’ve dropped my share of Lms throughout my career. But, in the example below, notice how the button copy elevates the mundane and creates momentum.
In this button, the user’s curiosity is both acknowledged and depicted. The wording is natural and evokes a confident, relaxed, informal tone. Perhaps users would like to see themselves as that kind of person (or they are that kind of person), so they’re tempted to tap the button.
The conversational AI revolution demands that talented writers step up and contribute clear, concise, upbeat copy. But bot design requires a certain temperament that few writers have. Writers can’t just go into their caves, emerge with a piece of work, hand it in and call it a day. Instead, they have to live inside the trenches of NLP solutions and be willing to review all transcripts and results and make continuous changes.
They have to deal in logic maps, intents, entities, follow-ups, fallbacks, training phrases, queries, default messages, error messages, and be willing and able to augment their abilities with machine learning tools.
When expert writers make peace with the tech, they can contribute to an AI’s evolving brain in a fluid, continuous way. That’s what’s needed for conversational AI to transcend the ordinary, for the betterment of business and humanity. Once writers get the hang of the tech and the lightbulb goes on, they may very well discover that writing for AI is more rewarding than writing for any other medium.