With the "Deus Ex Mach­ina? — Test­ing AI Tools"-series we want to show you dif­fer­ent tools, that aim to sim­pli­fy writ­ing, design and research by using Arti­fi­cial Intel­li­gence. More on the "Deus Ex Machina?"-series can be found here.

Overview

"In a soci­ety where time is pre­cious and inform­a­tion is every­where, we want to give our users pre­cise, cus­tom­ized and con­cise answers." This is the self-pro­claimed object­ive of the Per­plex­ity AI search engine.

What is spe­cial about this search engine is its appear­ance as an AI-sup­por­ted chat. You can think of it as a mix­ture of Chat­G­PT and Google. Ques­tions can be asked in nat­ur­al lan­guage and Per­plex­ity AI also responds in nat­ur­al lan­guage. To do this, it searches its data­base and the Inter­net and com­piles a short, con­cise text from the res­ults. This text is provided with ref­er­ences so that users can see which sources Per­plex­ity used to gen­er­ate the answers it provided. They also have the option of click­ing through to the source web­site them­selves and delving deep­er into the top­ic. In addi­tion, Per­plex­ity sug­gests fur­ther ques­tions after each search query, which can be used to research the top­ic in ques­tion even fur­ther. The col­lec­ted search quer­ies appear as a con­ver­sa­tion — sim­il­ar to Chat­G­PT — and can be saved and col­lec­ted. This allows users to access the res­ults again later.

In addi­tion to the gen­er­al search, users can also spe­cify their search para­met­ers and, for example, spe­cify that only sci­entif­ic papers, videos or Red­dit dis­cus­sions should be used as a data basis.

Per­plex­ity AI can be used free of charge and without regis­tra­tion via the browser. Per­plex­ity AI is also avail­able as an app for Android and IOS. When regis­ter­ing via email address, func­tions such as sav­ing search quer­ies and gen­er­at­ing images can be used. Per­plex­ity AI is ad-free, so res­ults are not sponsored. The Pro ver­sion of the search engine can be pur­chased for $20 per month or $200 per year. In addi­tion to unlim­ited quer­ies, bet­ter AI applic­a­tions and the abil­ity to fur­ther per­son­al­ize the search, Per­plex­ity AI also offers an AI-sup­por­ted assist­ant with its Co-Pilot in the Pro Ver­sion. It is designed to help deliv­er the best pos­sible res­ults — for example, by ask­ing users in detail what exactly they want to know.

The AI behind the application

With this new type of search engine, Per­plex­ity AI has entered the field of estab­lished search engines as an excit­ing new play­er. The com­pany was foun­ded in 2022 by Ara­vind Srinivas, Denis Yar­ats, Jonny Ho and Andy Kon­w­inski. In its short com­pany his­tory, the star­tup has already caused a stir with prom­in­ent investors such as Jeff Bezos, Susan Woj­cicki and Anderj Karparthy.

The stand­ard ver­sion of Per­plex­ity AI is based on Chat­G­PT 3.5 from OpenAI and com­bines this AI applic­a­tion with its own large lan­guage mod­el that works with nat­ur­al lan­guage pro­cessing. How­ever, the com­pany is keep­ing a lid on the exact func­tion­al­ity of its mod­el. The Pro ver­sion includes the new­er ver­sion GPT 4.0 instead of Chat­G­PT 3.5, as well as Claude‑2–1 and Gem­ini Pro AIs from oth­er developers.

Per­plex­ity AI is con­stantly being updated and expan­ded. As the com­pany is only at the begin­ning of its his­tory, it can be expec­ted that its own tech­no­logy will be fur­ther developed and new applic­a­tions launched in the upcom­ing years. Per­plex­ity AI util­izes user data to fur­ther train its AI, but the com­pany does not spe­cify which data is used for this pur­pose. This func­tion can be con­fined in the Pro version.

In June 2024, Per­plex­ity AI's work was cri­ti­cized on sev­er­al fronts. For example, Per­plex­ity is said to have repeatedly used and shared media con­tent that is hid­den behind a pay­wall — without the con­sent of the media in ques­tion. It has also been repor­ted that Per­plex­ity AI uses crawl­ers that viol­ate the so-called Robots Exclu­sion Stand­ard (robots.txt). With robots.txt, web­sites can par­tially or com­pletely block con­tent for crawl­ers. This is not a bind­ing law, but a kind of code of con­duct that reput­able crawl­ers adhere to. CEO Ara­vind Srinivas stated that the crawl­er used by Per­plex­ity is from a third-party pro­vider and not Perplexity's own crawl­er. How­ever, the com­pany did not provide any inform­a­tion about the con­sequences for the third-party pro­vider. Experts see the Per­plex­ity case as a good example of the prob­lems of cur­rent AI devel­op­ment, in which eth­ic­al stand­ards and the will to innov­ate must be con­stantly recon­ciled and weighed up against each other.

How­ever, in addi­tion to these copy­right issues, oth­er cri­ti­cisms were also voiced, arguing that Perplexity's answers too often relied on AI-gen­er­ated con­tent, some of which con­tained false inform­a­tion and was highly sus­cept­ible to so-called "hal­lu­cin­a­tions", also known as "bull­shit­ting". This is the phe­nomen­on that occurs when gen­er­at­ive AIs are asked for answers for which they can­not find suf­fi­cient mater­i­al in their data­bases and there­fore begin to fill the gaps with newly gen­er­ated inform­a­tion, which is often incor­rect. Accord­ing to research by For­bes magazine, Per­plex­ity AI uses AI-gen­er­ated sources for every third prompt on average.

The rhetorical potential of the tool

At first glance, Per­plex­ity AI's greatest poten­tial as a search engine lies in the inven­tio. It helps to find and pre­pare con­tent. How­ever, by out­put­ting the answers in nat­ur­al lan­guage, Per­plex­ity AI primar­ily facil­it­ates the trans­ition from inven­tio to eloc­u­tio, as the search engine already for­mu­lates the answers in a way that can be used in a text later on. This way, the applic­a­tion gives its users clear examples of how their own text can sound by the end and how the inform­a­tion can be edited. Of course, Per­plex­ity does not take the entire writ­ing pro­cess off the user's hands; the per­son behind the key­board still has to decide which inform­a­tion is ulti­mately relevant.

The pos­sib­il­ity of a focused search also gives users the oppor­tun­ity to search for suit­able inform­a­tion for dif­fer­ent tar­get groups and thus be able to speak to them more appro­pri­ately. How­ever, the tar­get group ori­ent­a­tion is also a major weak­ness of the tool. Per­plex­ity AI only provides inform­a­tion in a stand­ard­ized style that is as close as pos­sible to every­day lan­guage. Spe­cif­ic adapt­a­tions to oth­er lan­guage registers, as required by dif­fer­ent tar­get groups, can­not be made. The inform­a­tion provided would there­fore have to be reph­rased by the users, to be applic­able to spe­cif­ic media. For example, if the goal is to cre­ate a social media post, the response from Per­plex­ity AI must still be adap­ted to the highly abbre­vi­ated and spe­cif­ic con­tent form of social media platforms.

By cit­ing sources in the search res­ults, Per­plex­ity AI cre­ates more trust­wor­thi­ness and thus improves the per­suas­ive­ness of the con­tent of the text — known as logos in rhet­or­ic. The more logic­ally sound and cor­rectly worded a text is, the more con­vin­cing its con­tent will be. Accord­ing to its own inform­a­tion, the AI is trained to give pref­er­ence to cer­tain sources that are con­sidered par­tic­u­larly reli­able, such as the New York Times. How­ever, this applic­a­tion did not always work reli­ably in our test. Par­tic­u­larly for niche top­ics, res­ults from less reli­able sources such as Wiki­pe­dia were often dis­played. The cri­ti­cism of Per­plex­ity also high­lighted the fre­quent use of AI-gen­er­ated sources, which in turn are less reliable.

In gen­er­al, this inten­ded focus on recog­nized and estab­lished media and sources also cre­ates a cer­tain bias in the res­ults. The opin­ions and real­it­ies of life of minor­it­ies there­fore have less oppor­tun­ity to be seen. The author­ity of inter­pret­a­tion in the dis­course remains with famil­i­ar voices, which fur­ther rein­forces biases.

Per­plex­ity AI can not only out­put text, but also images. Pro account users also have the option of hav­ing match­ing images gen­er­ated for their search quer­ies. If users also want to integ­rate suit­able images into a text to make it more con­vin­cing, they do not have to use anoth­er tool to do so, but can cre­ate text and images that match each oth­er with the help of Per­plex­ity AI.

Usage in science communication

Per­plex­ity AI can be a good part­ner for sci­ence com­mu­nic­a­tion, as the tool com­bines vari­ous func­tions (inform­a­tion search, inform­a­tion pro­cessing, image search, image gen­er­a­tion) and thus saves users time. By focus­ing the search on purely aca­dem­ic sources, a high­er level of sci­entif­ic qual­ity can also be guar­an­teed here than with con­ven­tion­al search engines.

The AI behind Per­plex­ity AI is trained to provide answers that are as com­pre­hens­ible, pre­cise and con­cise as pos­sible. Para­met­ers that are also very import­ant for sci­entif­ic com­mu­nic­a­tion. How­ever, the tool's lack of flex­ib­il­ity is not to be under­es­tim­ated. It is hardly pos­sible to focus the text pro­duced on a tar­get audi­ence with Per­plex­ity AI, but this is essen­tial in sci­ence com­mu­nic­a­tion. Per­plex­ity AI does not dif­fer­en­ti­ate in its response as to wheth­er inform­a­tion should be under­stood by a sci­entif­ic audi­ence, inter­ested layper­sons or chil­dren. Users must then per­form this task them­selves in the cre­ation of sci­ence com­mu­nic­a­tion content.

The cri­ti­cism leveled against Per­plex­ity AI does not make its use in sci­ence com­mu­nic­a­tion advis­able. Reli­ab­il­ity and trust­wor­thi­ness are ele­ment­ary factors in sci­ence com­mu­nic­a­tion. The tool does not appear to be able to guar­an­tee either of these, which is why, des­pite its gen­er­ally good struc­ture, it should only be used with cau­tion and not without double check­ing the content.

Wrap-up

Per­plex­ity AI offers an excit­ing new way of using search engines and is very suit­able for every­day use. It saves its users time and scores points for being intu­it­ive and easy to under­stand. Even without regis­tra­tion or sub­scrip­tion, the ser­vice is help­ful and offers good fea­tures. By focus­ing on recog­nized media and sources, any exist­ing biases are repro­duced and rein­forced. Per­plex­ity AI is help­ful for the every­day work of (sci­ence) com­mu­nic­at­ors, but should be used with cau­tion in view of the tool's repeated "hal­lu­cin­a­tions". In gen­er­al, Per­plex­ity AI is not a secret magic-tool for sci­ence com­mu­nic­a­tion: its tar­get group ori­ent­a­tion can­not be finely adjus­ted enough for that. How­ever, it makes the ini­tial steps easi­er, shows excit­ing per­spect­ives for fur­ther thought pro­cesses and can save users time and clicks by sum­mar­iz­ing search res­ults. Wheth­er users want to sup­port Perplexity's way of work­ing, which can some­times be clas­si­fied as uneth­ic­al, is ulti­mately up to the respect­ive users to decide.