How open-source AI could modernise public services

Arti­fi­cial intel­li­gence (AI), and more spe­cific­ally large lan­guage mod­els (LLMs) rep­res­ent a major oppor­tun­ity to trans­form pub­lic ser­vices. AI can be used in many areas to improve effi­ciency, the qual­ity of ser­vices provided to cit­izens and decision-mak­ing. How­ever, the imple­ment­a­tion of AI in pub­lic ser­vices presents major chal­lenges. First, the chosen solu­tion must guar­an­tee fair treat­ment and trans­par­ency of decisions and actions on a case or ensure respect for fun­da­ment­al rights through­out its use. In addi­tion, the rig­or­ous pro­tec­tion of per­son­al data, which is often sens­it­ive in the con­text of pub­lic ser­vices, is a sig­ni­fic­ant secur­ity issue. Finally, the trans­par­ency of decisions is a major factor in the trust placed in the solu­tions used and their accept­ab­il­ity to cit­izens. Thus, the use of a solu­tion offer­ing a high level of trans­par­ency is an asset in the imple­ment­a­tion and accept­ance of arti­fi­cial intel­li­gence solu­tions. But in view of the com­plex­ity of the sub­ject, the cri­ter­ia for ensur­ing the expec­ted level of trans­par­ency are far from easy to define.

The major lan­guage mod­els are based on neur­al net­works trained on a very large amount of data. From a sequence of words, they stat­ist­ic­ally determ­ine the word that best matches the giv­en sequence. By apply­ing this prin­ciple recurs­ively, LLMs are able to pro­duce struc­tured texts, giv­ing the impres­sion that the machine is ana­lys­ing and under­stand­ing the ques­tion asked.

The text pro­duced will there­fore depend on:

• the algorithms used, which will enable the mod­el to weigh the import­ance of each word in a sen­tence in rela­tion to the oth­ers. This capa­city is provided in par­tic­u­lar through “trans­former” type archi­tec­tures¹.
• the weight assigned to the dif­fer­ent neur­ons, which will enable the net­work to be activ­ated in order to pro­duce the out­put data;
• the learn­ing cor­pus, which has a dir­ect impact on the determ­in­a­tion of the weights used by the model.

The 4 prin­ciples (use, study, modi­fy, share) asso­ci­ated with free soft­ware² must there­fore be applied to all these ele­ments³. This is still a sub­ject of debate and is there­fore a source of much con­fu­sion⁴. For example, some AIs claim­ing to be free have usage restric­tions that go against the defined prin­ciples⁵. After a long pro­cess, the Open Source Ini­ti­at­ive (OSI), which brings togeth­er research­ers, law­yers, policy makers, act­iv­ists and rep­res­ent­at­ives of large tech­no­logy com­pan­ies, has pro­posed a defin­i­tion that cor­rel­ates the four freedoms asso­ci­ated with free soft­ware and the ele­ments on which machine learn­ing mod­els are based.

Accord­ing to the Open Source Ini­ti­at­ive, a free machine learn­ing sys­tem must include the fol­low­ing ele­ments⁶:

• suf­fi­ciently detailed inform­a­tion on the data used to train the sys­tem, enabling a com­pet­ent per­son to build a sub­stan­tially equi­val­ent sys­tem. This inform­a­tion must be avail­able under terms approved by the OSI;
• the source code of the AI, includ­ing the infer­ence code to execute the model;
• all the learned para­met­ers that are super­im­posed on the mod­el archi­tec­ture to pro­duce an out­put from a giv­en input.

The pub­lic­a­tion of the learn­ing cor­pus is there­fore not com­puls­ory, but a detailed descrip­tion of the lat­ter must be included. It is clear that many mod­els offer­ing excel­lent per­form­ance and describ­ing them­selves as open source do not com­ply with this last point. These are referred to as open-weight mod­els. A com­par­is­on of AI mod­els is also avail­able from the Pôle d’Ex­pert­ise de la Régu­la­tion Numérique (PER­eN).

The source code is human-read­able and provides access to the algorithms used. The weights are the res­ult of train­ing and rep­res­ent the know­ledge of the mod­el. In the case of open-weight mod­els, this know­ledge can be cus­tom­ised through a fine-tun­ing pro­cess⁷.

How­ever, this does not allow for total trans­par­ency, such as the detec­tion of bias or “pois­on­ing” attacks, which con­sist of alter­ing the know­ledge of a mod­el without these modi­fic­a­tions being eas­ily detect­able by stand­ard tests⁸⁹. Only a free mod­el that provides access to its learn­ing cor­pus guar­an­tees a total level of trans­par­ency, in par­tic­u­lar by allow­ing com­plete con­trol of its train­ing. How­ever, this approach of recon­struct­ing from the sources still requires sig­ni­fic­ant com­put­ing resources that few entit­ies are able to acquire.

On 30^th Octo­ber 2023, Pres­id­ent Biden issued an exec­ut­ive order entitled Safe, Secure, and Trust­worthy Devel­op­ment and Use of Arti­fi­cial Intel­li­gence to assess the risks and bene­fits of found­a­tion mod­els for which weights are avail­able. The report res­ult­ing from this study¹⁰ recog­nises the bene­fits of open access to mod­el weights, such as innov­a­tion and research, but also high­lights the poten­tial risks, includ­ing the pos­sib­il­ity of mali­cious use, the remov­al of safety mech­an­isms and the impact on com­pet­i­tion. The report con­cludes that the cur­rent data is insuf­fi­cient to defin­it­ively determ­ine wheth­er restric­tions on open-weight mod­els are jus­ti­fied and recom­mends act­ive mon­it­or­ing of these models.

Closed mod­els, even if they do not bene­fit from the same level of trans­par­ency and adapt­ab­il­ity as their free or open-weight coun­ter­parts, are not without advant­ages. They are less sub­ject to the risks of manip­u­la­tion men­tioned above because their weights can­not be mod­i­fied by a third party, the risks to the intel­lec­tu­al prop­erty of the train­ing data are borne by the mod­el sup­pli­er, the pub­lish­er can quickly act on its mod­el in order to react in the event of abuse, thus help­ing to mit­ig­ate the poten­tial risks asso­ci­ated with AI, such as the dis­sem­in­a­tion of inap­pro­pri­ate con­tent¹¹. How­ever, all this is at the expense of the autonomy that can be had over the AI model.

The use of open AIs as defined by the OSI has many advant­ages. First of all, the trans­par­ency of their oper­a­tion is guar­an­teed, since it is pos­sible to dir­ectly access and modi­fy their source code and inspect the train­ing data.

This pos­sib­il­ity is a fun­da­ment­al guar­an­tee since each mod­el used can be sub­jec­ted to in-depth veri­fic­a­tion to ensure that the decision-mak­ing pro­cess com­plies with cur­rent legis­la­tion and does not present any dis­crim­in­at­ory bias, for example. On the oth­er hand, when AI is used in the con­text of retriev­al-aug­men­ted gen­er­a­tion (RAG¹²), the level of trans­par­ency that must be required may be lower because the data used to for­mu­late the responses are provided through an algorithm over which it is easi­er to have the expec­ted level of con­trol. As the body of answers is provided by con­ven­tion­al search algorithms, it is rel­at­ively easy to provide the end user with the expec­ted answer, the raw data and their level of con­fid­ence. How­ever, this requires a crit­ic­al eye on the part of the end user.

Even if the State’s mis­sions are rel­at­ively spe­cif­ic by their very nature, there are many cases of use that are sim­il­ar to those that can be found in private com­pan­ies, namely provid­ing an answer to a ques­tion by using a cor­pus of doc­u­ments with the help of clas­sic or vec­tor search algorithms based on the concept of sim­il­ar­ity¹³. It is there­fore not sur­pris­ing to see a con­ver­gence in the mod­els used in both worlds. For the State, the decid­ing cri­terion in the choice of mod­els will there­fore be related to the pre­ser­va­tion of per­son­al inform­a­tion or sens­it­ive inform­a­tion trans­mit­ted to AI models.

The use of open-source solu­tions makes it pos­sible to drastic­ally reduce expenses

Bey­ond the aspects men­tioned above, the use of open-source solu­tions also allows the State to dis­sem­in­ate its work so that it can be reused by the pub­lic or private sec­tor. For example, the DGFiP has pub­lished work on a mod­el for syn­thes­ising par­lia­ment­ary amend­ments¹⁴¹⁵. They are thus able to act­ively share their know­ledge with­in the lim­its of con­fid­en­ti­al­ity neces­sary for sov­er­eign missions.

Finally, the use of open-source solu­tions makes it pos­sible to drastic­ally reduce expenses, by lim­it­ing them to tech­nic­al sup­port without licence costs.

The use of AI under a free licence also presents vari­ous chal­lenges. First of all, the imple­ment­a­tion of free solu­tions requires a good under­stand­ing of how the under­ly­ing mod­els work. In addi­tion to this com­plex­ity, there is also the need for tech­nic­al skills to adapt the mod­els to busi­ness needs, to have the data neces­sary for learn­ing, to con­fig­ure the mod­el (fine-tun­ing), if the busi­ness applic­a­tion requires it, to deploy it in the admin­is­tra­tion’s IS and to guar­an­tee the highest level of security.

In addi­tion, their ongo­ing main­ten­ance and cor­rect­ive main­ten­ance require a sig­ni­fic­ant invest­ment of time, both to update the mod­els or ensure a sat­is­fact­ory level of non-regres­sion and to ensure that they func­tion prop­erly. Although the code is free, the use of these AIs often also requires IT infra­struc­tures based on spe­cial­ised com­put­ing units, which can rep­res­ent an indir­ect cost. Finally, the qual­ity of open-source mod­els can vary con­sid­er­ably, par­tic­u­larly depend­ing on the busi­ness cases to be addressed, and there are no abso­lute guar­an­tees as to their per­form­ance. It is there­fore essen­tial to define the expect­a­tions pre­cisely with the busi­ness teams and to veri­fy the expec­ted res­ults before any ver­sion is put into service.

The integ­ra­tion of arti­fi­cial intel­li­gence with­in pub­lic ser­vices rep­res­ents a unique oppor­tun­ity to improve the effi­ciency and qual­ity of ser­vices provided to cit­izens and decision-mak­ing in a con­text of strain on avail­able human resources. Free lan­guage mod­els seem to be par­tic­u­larly well-suited to this challenge.

Des­pite the chal­lenges, the advant­ages of free AIs are numer­ous. They pro­mote innov­a­tion, reduce costs and strengthen the autonomy of administrations.

How­ever, it is essen­tial to study in depth the eth­ic­al issues related to the use of AI in the pub­lic sec­tor. It is neces­sary to put in place pro­cesses and meth­ods to guard against algorithmic bias and to guar­an­tee the reas­on­able use of tech­no­lo­gies, ensur­ing that they are mon­itored by digit­al and leg­al experts, or even by cit­izens themselves.

Dis­claim­er: The con­tent of this art­icle is the sole respons­ib­il­ity of its authors and has no scope oth­er than that of inform­a­tion and aca­dem­ic research.